11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 441da177e4SLinus Torvalds 451da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 46ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 471da177e4SLinus Torvalds 4815d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4915d99e02SRick Jones * interpret the window field as a signed quantity. 5015d99e02SRick Jones */ 51ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5215d99e02SRick Jones 5346d3ceabSEric Dumazet /* Default TSQ limit of two TSO segments */ 5446d3ceabSEric Dumazet int sysctl_tcp_limit_output_bytes __read_mostly = 131072; 5546d3ceabSEric Dumazet 561da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 571da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 581da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 591da177e4SLinus Torvalds */ 60ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 611da177e4SLinus Torvalds 62ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 6397b1ce25SShan Wei int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; 645d424d5aSJohn Heffner 6535089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 66ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6735089bb2SDavid S. Miller 68c9bee3b7SEric Dumazet unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; 69c9bee3b7SEric Dumazet EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); 70c9bee3b7SEric Dumazet 7146d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 7246d3ceabSEric Dumazet int push_one, gfp_t gfp); 73519855c5SWilliam Allen Simpson 7467edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 75cf533ea5SEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 766ff03ac3SIlpo Järvinen { 776ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 786ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7966f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 809e412ba7SIlpo Järvinen 81fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 821da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 838512430eSIlpo Järvinen 8466f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 856ba8a3b1SNandita Dukkipati if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 866a5dc9e5SEric Dumazet icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 87750ea2baSYuchung Cheng tcp_rearm_rto(sk); 881da177e4SLinus Torvalds } 89f19c29e3SYuchung Cheng 90f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 91f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 926a5dc9e5SEric Dumazet } 931da177e4SLinus Torvalds 941da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 951da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 961da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 971da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 981da177e4SLinus Torvalds * invalid. OK, let's make this for now: 991da177e4SLinus Torvalds */ 100cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 1011da177e4SLinus Torvalds { 102cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1039e412ba7SIlpo Järvinen 10490840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 1051da177e4SLinus Torvalds return tp->snd_nxt; 1061da177e4SLinus Torvalds else 10790840defSIlpo Järvinen return tcp_wnd_end(tp); 1081da177e4SLinus Torvalds } 1091da177e4SLinus Torvalds 1101da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1111da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1121da177e4SLinus Torvalds * 1131da177e4SLinus Torvalds * 1. It is independent of path mtu. 1141da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1151da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1161da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1171da177e4SLinus Torvalds * large MSS. 1181da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1191da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1201da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1211da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1221da177e4SLinus Torvalds * probably even Jumbo". 1231da177e4SLinus Torvalds */ 1241da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1251da177e4SLinus Torvalds { 1261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 127cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1281da177e4SLinus Torvalds int mss = tp->advmss; 1291da177e4SLinus Torvalds 1300dbaee3bSDavid S. Miller if (dst) { 1310dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1320dbaee3bSDavid S. Miller 1330dbaee3bSDavid S. Miller if (metric < mss) { 1340dbaee3bSDavid S. Miller mss = metric; 1351da177e4SLinus Torvalds tp->advmss = mss; 1361da177e4SLinus Torvalds } 1370dbaee3bSDavid S. Miller } 1381da177e4SLinus Torvalds 1391da177e4SLinus Torvalds return (__u16)mss; 1401da177e4SLinus Torvalds } 1411da177e4SLinus Torvalds 1421da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1431da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 144cf533ea5SEric Dumazet static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 1451da177e4SLinus Torvalds { 146463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1471da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1481da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1491da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1501da177e4SLinus Torvalds 1516687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1521da177e4SLinus Torvalds 1536687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1541da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1551da177e4SLinus Torvalds 156463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1571da177e4SLinus Torvalds cwnd >>= 1; 1581da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1591da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1601da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1611da177e4SLinus Torvalds } 1621da177e4SLinus Torvalds 16367edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 16440efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 165cf533ea5SEric Dumazet struct sock *sk) 1661da177e4SLinus Torvalds { 167463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 168463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 169bcefe17cSCong Wang const struct dst_entry *dst = __sk_dst_get(sk); 1701da177e4SLinus Torvalds 17135089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 17235089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 173463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1741da177e4SLinus Torvalds 1751da177e4SLinus Torvalds tp->lsndtime = now; 1761da177e4SLinus Torvalds 1771da177e4SLinus Torvalds /* If it is a reply for ato after last received 1781da177e4SLinus Torvalds * packet, enter pingpong mode. 1791da177e4SLinus Torvalds */ 180bcefe17cSCong Wang if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato && 181bcefe17cSCong Wang (!dst || !dst_metric(dst, RTAX_QUICKACK))) 182463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1831da177e4SLinus Torvalds } 1841da177e4SLinus Torvalds 18567edfef7SAndi Kleen /* Account for an ACK we sent. */ 18640efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1871da177e4SLinus Torvalds { 188463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 189463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1901da177e4SLinus Torvalds } 1911da177e4SLinus Torvalds 19285f16525SYuchung Cheng 19385f16525SYuchung Cheng u32 tcp_default_init_rwnd(u32 mss) 19485f16525SYuchung Cheng { 19585f16525SYuchung Cheng /* Initial receive window should be twice of TCP_INIT_CWND to 1969ef71e0cSWeiping Pan * enable proper sending of new unsent data during fast recovery 19785f16525SYuchung Cheng * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 19885f16525SYuchung Cheng * limit when mss is larger than 1460. 19985f16525SYuchung Cheng */ 20085f16525SYuchung Cheng u32 init_rwnd = TCP_INIT_CWND * 2; 20185f16525SYuchung Cheng 20285f16525SYuchung Cheng if (mss > 1460) 20385f16525SYuchung Cheng init_rwnd = max((1460 * init_rwnd) / mss, 2U); 20485f16525SYuchung Cheng return init_rwnd; 20585f16525SYuchung Cheng } 20685f16525SYuchung Cheng 2071da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2081da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2091da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2101da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2111da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2121da177e4SLinus Torvalds * This MUST be enforced by all callers. 2131da177e4SLinus Torvalds */ 2141da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 2151da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 21631d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21731d12926Slaurent chavey __u32 init_rcv_wnd) 2181da177e4SLinus Torvalds { 2191da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2201da177e4SLinus Torvalds 2211da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2221da177e4SLinus Torvalds if (*window_clamp == 0) 2231da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 2241da177e4SLinus Torvalds space = min(*window_clamp, space); 2251da177e4SLinus Torvalds 2261da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2271da177e4SLinus Torvalds if (space > mss) 2281da177e4SLinus Torvalds space = (space / mss) * mss; 2291da177e4SLinus Torvalds 2301da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 23115d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 23215d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 23315d99e02SRick Jones * we will truncate our initial window offering to 32K-1 23415d99e02SRick Jones * unless the remote has sent us a window scaling option, 23515d99e02SRick Jones * which we interpret as a sign the remote TCP is not 23615d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2371da177e4SLinus Torvalds */ 23815d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2391da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 24015d99e02SRick Jones else 24115d99e02SRick Jones (*rcv_wnd) = space; 24215d99e02SRick Jones 2431da177e4SLinus Torvalds (*rcv_wscale) = 0; 2441da177e4SLinus Torvalds if (wscale_ok) { 2451da177e4SLinus Torvalds /* Set window scaling on max possible window 2461da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2471da177e4SLinus Torvalds */ 2481da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 249316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2501da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2511da177e4SLinus Torvalds space >>= 1; 2521da177e4SLinus Torvalds (*rcv_wscale)++; 2531da177e4SLinus Torvalds } 2541da177e4SLinus Torvalds } 2551da177e4SLinus Torvalds 2561da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 25785f16525SYuchung Cheng if (!init_rcv_wnd) /* Use default unless specified otherwise */ 25885f16525SYuchung Cheng init_rcv_wnd = tcp_default_init_rwnd(mss); 259b1afde60SNandita Dukkipati *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 2601da177e4SLinus Torvalds } 2611da177e4SLinus Torvalds 2621da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2631da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2641da177e4SLinus Torvalds } 2654bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2661da177e4SLinus Torvalds 2671da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2681da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2691da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2701da177e4SLinus Torvalds * frame. 2711da177e4SLinus Torvalds */ 27240efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2731da177e4SLinus Torvalds { 2741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2758e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2761da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2771da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2781da177e4SLinus Torvalds 2791da177e4SLinus Torvalds /* Never shrink the offered window */ 2801da177e4SLinus Torvalds if (new_win < cur_win) { 2811da177e4SLinus Torvalds /* Danger Will Robinson! 2821da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2831da177e4SLinus Torvalds * we will not be able to advertise a zero 2841da177e4SLinus Torvalds * window in time. --DaveM 2851da177e4SLinus Torvalds * 2861da177e4SLinus Torvalds * Relax Will Robinson. 2871da177e4SLinus Torvalds */ 2888e165e20SFlorian Westphal if (new_win == 0) 2898e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2908e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 291607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2921da177e4SLinus Torvalds } 2931da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2941da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2951da177e4SLinus Torvalds 2961da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2971da177e4SLinus Torvalds * scaled window. 2981da177e4SLinus Torvalds */ 29915d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 3001da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 3011da177e4SLinus Torvalds else 3021da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 3031da177e4SLinus Torvalds 3041da177e4SLinus Torvalds /* RFC1323 scaling applied */ 3051da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 3061da177e4SLinus Torvalds 3071da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 3088e165e20SFlorian Westphal if (new_win == 0) { 3091da177e4SLinus Torvalds tp->pred_flags = 0; 3108e165e20SFlorian Westphal if (old_win) 3118e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 3128e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3138e165e20SFlorian Westphal } else if (old_win == 0) { 3148e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3158e165e20SFlorian Westphal } 3161da177e4SLinus Torvalds 3171da177e4SLinus Torvalds return new_win; 3181da177e4SLinus Torvalds } 3191da177e4SLinus Torvalds 32067edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 321*735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 322bdf1ee5dSIlpo Järvinen { 32330e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 32430e502a3SDaniel Borkmann 3254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 326bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3274de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 32830e502a3SDaniel Borkmann else if (tcp_ca_needs_ecn(sk)) 32930e502a3SDaniel Borkmann INET_ECN_xmit(sk); 330bdf1ee5dSIlpo Järvinen } 331bdf1ee5dSIlpo Järvinen 33267edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 333*735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 334bdf1ee5dSIlpo Järvinen { 335bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 336bdf1ee5dSIlpo Järvinen 337bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 33830e502a3SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 33930e502a3SDaniel Borkmann tcp_ca_needs_ecn(sk)) { 3404de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 341bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34230e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 34330e502a3SDaniel Borkmann INET_ECN_xmit(sk); 344bdf1ee5dSIlpo Järvinen } 345bdf1ee5dSIlpo Järvinen } 346bdf1ee5dSIlpo Järvinen 347*735d3831SFlorian Westphal static void 348*735d3831SFlorian Westphal tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th, 34930e502a3SDaniel Borkmann struct sock *sk) 350bdf1ee5dSIlpo Järvinen { 35130e502a3SDaniel Borkmann if (inet_rsk(req)->ecn_ok) { 352bdf1ee5dSIlpo Järvinen th->ece = 1; 35330e502a3SDaniel Borkmann if (tcp_ca_needs_ecn(sk)) 35430e502a3SDaniel Borkmann INET_ECN_xmit(sk); 35530e502a3SDaniel Borkmann } 356bdf1ee5dSIlpo Järvinen } 357bdf1ee5dSIlpo Järvinen 35867edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 35967edfef7SAndi Kleen * be sent. 36067edfef7SAndi Kleen */ 361*735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 362bdf1ee5dSIlpo Järvinen int tcp_header_len) 363bdf1ee5dSIlpo Järvinen { 364bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 365bdf1ee5dSIlpo Järvinen 366bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 367bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 368bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 369bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 370bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 371bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 372bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 373bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 374bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 375bdf1ee5dSIlpo Järvinen } 37630e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 377bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 378bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 379bdf1ee5dSIlpo Järvinen } 380bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 381bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 382bdf1ee5dSIlpo Järvinen } 383bdf1ee5dSIlpo Järvinen } 384bdf1ee5dSIlpo Järvinen 385e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 386e870a8efSIlpo Järvinen * auto increment end seqno. 387e870a8efSIlpo Järvinen */ 388e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 389e870a8efSIlpo Järvinen { 3907b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 3917b7fc97aSEric Dumazet 3922e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 393e870a8efSIlpo Järvinen skb->csum = 0; 394e870a8efSIlpo Järvinen 3954de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 396e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 397e870a8efSIlpo Järvinen 398cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 3997b7fc97aSEric Dumazet shinfo->gso_size = 0; 4007b7fc97aSEric Dumazet shinfo->gso_type = 0; 401e870a8efSIlpo Järvinen 402e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 403a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 404e870a8efSIlpo Järvinen seq++; 405e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 406e870a8efSIlpo Järvinen } 407e870a8efSIlpo Järvinen 408a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 40933f5f57eSIlpo Järvinen { 41033f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 41133f5f57eSIlpo Järvinen } 41233f5f57eSIlpo Järvinen 41333ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41433ad798cSAdam Langley #define OPTION_TS (1 << 1) 41533ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41689e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4172100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 41833ad798cSAdam Langley 41933ad798cSAdam Langley struct tcp_out_options { 4202100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4212100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 42233ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 42333ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 424bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 425bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4262100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4272100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 42833ad798cSAdam Langley }; 42933ad798cSAdam Langley 43067edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 43167edfef7SAndi Kleen * 43267edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 433fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 434fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4358e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 436fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 437fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 438fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 439fd6149d3SIlpo Järvinen * 440fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 441fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 442fd6149d3SIlpo Järvinen */ 44333ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 444bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 445bd0388aeSWilliam Allen Simpson { 4462100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 447bd0388aeSWilliam Allen Simpson 448bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4491a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4501a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 451bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 452bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45333ad798cSAdam Langley ptr += 4; 45433ad798cSAdam Langley } 45533ad798cSAdam Langley 456fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 457fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 458fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 459fd6149d3SIlpo Järvinen opts->mss); 460fd6149d3SIlpo Järvinen } 461fd6149d3SIlpo Järvinen 462bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 463bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46533ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 46633ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 46733ad798cSAdam Langley TCPOLEN_TIMESTAMP); 468bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 46933ad798cSAdam Langley } else { 470496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47140efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47240efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47340efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47440efc6faSStephen Hemminger } 47533ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 47633ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 47733ad798cSAdam Langley } 47833ad798cSAdam Langley 479bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 48033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48133ad798cSAdam Langley (TCPOPT_NOP << 16) | 48233ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 48333ad798cSAdam Langley TCPOLEN_SACK_PERM); 48433ad798cSAdam Langley } 48533ad798cSAdam Langley 486bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 48733ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 48833ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 48933ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 49033ad798cSAdam Langley opts->ws); 49133ad798cSAdam Langley } 49233ad798cSAdam Langley 49333ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 49433ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 49533ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 49640efc6faSStephen Hemminger int this_sack; 49740efc6faSStephen Hemminger 49840efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 49940efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 50040efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 50133ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 50240efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5032de979bdSStephen Hemminger 50433ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 50533ad798cSAdam Langley ++this_sack) { 50640efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 50740efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 50840efc6faSStephen Hemminger } 5092de979bdSStephen Hemminger 51040efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 51140efc6faSStephen Hemminger } 5122100c8d2SYuchung Cheng 5132100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5142100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5152100c8d2SYuchung Cheng 5162100c8d2SYuchung Cheng *ptr++ = htonl((TCPOPT_EXP << 24) | 5172100c8d2SYuchung Cheng ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 5182100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5192100c8d2SYuchung Cheng 5202100c8d2SYuchung Cheng memcpy(ptr, foc->val, foc->len); 5212100c8d2SYuchung Cheng if ((foc->len & 3) == 2) { 5222100c8d2SYuchung Cheng u8 *align = ((u8 *)ptr) + foc->len; 5232100c8d2SYuchung Cheng align[0] = align[1] = TCPOPT_NOP; 5242100c8d2SYuchung Cheng } 5252100c8d2SYuchung Cheng ptr += (foc->len + 3) >> 2; 5262100c8d2SYuchung Cheng } 52740efc6faSStephen Hemminger } 52840efc6faSStephen Hemminger 52967edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 53067edfef7SAndi Kleen * network wire format yet. 53167edfef7SAndi Kleen */ 53295c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 53333ad798cSAdam Langley struct tcp_out_options *opts, 534cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 535cf533ea5SEric Dumazet { 53633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 53795c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 538783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 53933ad798cSAdam Langley 540cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 54133ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 54233ad798cSAdam Langley if (*md5) { 54333ad798cSAdam Langley opts->options |= OPTION_MD5; 544bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 545cfb6eeb4SYOSHIFUJI Hideaki } 54633ad798cSAdam Langley #else 54733ad798cSAdam Langley *md5 = NULL; 548cfb6eeb4SYOSHIFUJI Hideaki #endif 54933ad798cSAdam Langley 55033ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 55133ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 55233ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 55333ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 55433ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 55533ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 55633ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 55733ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 55833ad798cSAdam Langley * going out. */ 55933ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 560bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 56133ad798cSAdam Langley 562bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 56333ad798cSAdam Langley opts->options |= OPTION_TS; 5647faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 56533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 566bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 56733ad798cSAdam Langley } 568bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 56933ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 57089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 571bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 57233ad798cSAdam Langley } 573bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 57433ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 575b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 576bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 57733ad798cSAdam Langley } 57833ad798cSAdam Langley 579783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 580783237e8SYuchung Cheng u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 581783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 582783237e8SYuchung Cheng if (remaining >= need) { 583783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 584783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 585783237e8SYuchung Cheng remaining -= need; 586783237e8SYuchung Cheng tp->syn_fastopen = 1; 587783237e8SYuchung Cheng } 588783237e8SYuchung Cheng } 589bd0388aeSWilliam Allen Simpson 590bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 59133ad798cSAdam Langley } 59233ad798cSAdam Langley 59367edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 59495c96174SEric Dumazet static unsigned int tcp_synack_options(struct sock *sk, 59533ad798cSAdam Langley struct request_sock *req, 59695c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 59733ad798cSAdam Langley struct tcp_out_options *opts, 5984957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 5998336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6004957faadSWilliam Allen Simpson { 60133ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 60295c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 60333ad798cSAdam Langley 60433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 60533ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 60633ad798cSAdam Langley if (*md5) { 60733ad798cSAdam Langley opts->options |= OPTION_MD5; 6084957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6094957faadSWilliam Allen Simpson 6104957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6114957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6124957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6134957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6144957faadSWilliam Allen Simpson */ 615de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 61633ad798cSAdam Langley } 61733ad798cSAdam Langley #else 61833ad798cSAdam Langley *md5 = NULL; 61933ad798cSAdam Langley #endif 62033ad798cSAdam Langley 6214957faadSWilliam Allen Simpson /* We always send an MSS option. */ 62233ad798cSAdam Langley opts->mss = mss; 6234957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 62433ad798cSAdam Langley 62533ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 62633ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 62789e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6284957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 62933ad798cSAdam Langley } 630de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 63133ad798cSAdam Langley opts->options |= OPTION_TS; 6327faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb); 63333ad798cSAdam Langley opts->tsecr = req->ts_recent; 6344957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 63533ad798cSAdam Langley } 63633ad798cSAdam Langley if (likely(ireq->sack_ok)) { 63733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 638de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6394957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 64033ad798cSAdam Langley } 64189278c9dSYuchung Cheng if (foc != NULL && foc->len >= 0) { 6428336886fSJerry Chu u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 6438336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 6448336886fSJerry Chu if (remaining >= need) { 6458336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 6468336886fSJerry Chu opts->fastopen_cookie = foc; 6478336886fSJerry Chu remaining -= need; 6488336886fSJerry Chu } 6498336886fSJerry Chu } 6504957faadSWilliam Allen Simpson 6514957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 65233ad798cSAdam Langley } 65333ad798cSAdam Langley 65467edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 65567edfef7SAndi Kleen * final wire format yet. 65667edfef7SAndi Kleen */ 65795c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 65833ad798cSAdam Langley struct tcp_out_options *opts, 659cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 660cf533ea5SEric Dumazet { 66133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 66295c96174SEric Dumazet unsigned int size = 0; 663cabeccbdSIlpo Järvinen unsigned int eff_sacks; 66433ad798cSAdam Langley 6655843ef42SAndi Kleen opts->options = 0; 6665843ef42SAndi Kleen 66733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 66833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 66933ad798cSAdam Langley if (unlikely(*md5)) { 67033ad798cSAdam Langley opts->options |= OPTION_MD5; 67133ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 67233ad798cSAdam Langley } 67333ad798cSAdam Langley #else 67433ad798cSAdam Langley *md5 = NULL; 67533ad798cSAdam Langley #endif 67633ad798cSAdam Langley 67733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 67833ad798cSAdam Langley opts->options |= OPTION_TS; 6797faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 68033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 68133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 68233ad798cSAdam Langley } 68333ad798cSAdam Langley 684cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 685cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 68695c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 68733ad798cSAdam Langley opts->num_sack_blocks = 68895c96174SEric Dumazet min_t(unsigned int, eff_sacks, 68933ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 69033ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 69133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 69233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 69333ad798cSAdam Langley } 69433ad798cSAdam Langley 69533ad798cSAdam Langley return size; 69640efc6faSStephen Hemminger } 6971da177e4SLinus Torvalds 69846d3ceabSEric Dumazet 69946d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 70046d3ceabSEric Dumazet * 70146d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 70246d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 70346d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 70446d3ceabSEric Dumazet * 70546d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 70646d3ceabSEric Dumazet * needs to be reallocated in a driver. 7078e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 70846d3ceabSEric Dumazet * 70946d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 71046d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 71146d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 71246d3ceabSEric Dumazet */ 71346d3ceabSEric Dumazet struct tsq_tasklet { 71446d3ceabSEric Dumazet struct tasklet_struct tasklet; 71546d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 71646d3ceabSEric Dumazet }; 71746d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 71846d3ceabSEric Dumazet 7196f458dfbSEric Dumazet static void tcp_tsq_handler(struct sock *sk) 7206f458dfbSEric Dumazet { 7216f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7226f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 7236f458dfbSEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 724bf06200eSJohn Ogness tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 725bf06200eSJohn Ogness 0, GFP_ATOMIC); 7266f458dfbSEric Dumazet } 72746d3ceabSEric Dumazet /* 7288e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 72946d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 7308e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 73146d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 73246d3ceabSEric Dumazet */ 73346d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 73446d3ceabSEric Dumazet { 73546d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 73646d3ceabSEric Dumazet LIST_HEAD(list); 73746d3ceabSEric Dumazet unsigned long flags; 73846d3ceabSEric Dumazet struct list_head *q, *n; 73946d3ceabSEric Dumazet struct tcp_sock *tp; 74046d3ceabSEric Dumazet struct sock *sk; 74146d3ceabSEric Dumazet 74246d3ceabSEric Dumazet local_irq_save(flags); 74346d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 74446d3ceabSEric Dumazet local_irq_restore(flags); 74546d3ceabSEric Dumazet 74646d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 74746d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 74846d3ceabSEric Dumazet list_del(&tp->tsq_node); 74946d3ceabSEric Dumazet 75046d3ceabSEric Dumazet sk = (struct sock *)tp; 75146d3ceabSEric Dumazet bh_lock_sock(sk); 75246d3ceabSEric Dumazet 75346d3ceabSEric Dumazet if (!sock_owned_by_user(sk)) { 7546f458dfbSEric Dumazet tcp_tsq_handler(sk); 75546d3ceabSEric Dumazet } else { 75646d3ceabSEric Dumazet /* defer the work to tcp_release_cb() */ 7576f458dfbSEric Dumazet set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 75846d3ceabSEric Dumazet } 75946d3ceabSEric Dumazet bh_unlock_sock(sk); 76046d3ceabSEric Dumazet 76146d3ceabSEric Dumazet clear_bit(TSQ_QUEUED, &tp->tsq_flags); 76246d3ceabSEric Dumazet sk_free(sk); 76346d3ceabSEric Dumazet } 76446d3ceabSEric Dumazet } 76546d3ceabSEric Dumazet 7666f458dfbSEric Dumazet #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 7676f458dfbSEric Dumazet (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 768563d34d0SEric Dumazet (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 769563d34d0SEric Dumazet (1UL << TCP_MTU_REDUCED_DEFERRED)) 77046d3ceabSEric Dumazet /** 77146d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 77246d3ceabSEric Dumazet * @sk: socket 77346d3ceabSEric Dumazet * 77446d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 77546d3ceabSEric Dumazet * actions before socket release. 77646d3ceabSEric Dumazet */ 77746d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 77846d3ceabSEric Dumazet { 77946d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 7806f458dfbSEric Dumazet unsigned long flags, nflags; 78146d3ceabSEric Dumazet 7826f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 7836f458dfbSEric Dumazet do { 7846f458dfbSEric Dumazet flags = tp->tsq_flags; 7856f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 7866f458dfbSEric Dumazet return; 7876f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 7886f458dfbSEric Dumazet } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 7896f458dfbSEric Dumazet 7906f458dfbSEric Dumazet if (flags & (1UL << TCP_TSQ_DEFERRED)) 7916f458dfbSEric Dumazet tcp_tsq_handler(sk); 7926f458dfbSEric Dumazet 793c3f9b018SEric Dumazet /* Here begins the tricky part : 794c3f9b018SEric Dumazet * We are called from release_sock() with : 795c3f9b018SEric Dumazet * 1) BH disabled 796c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 797c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 798c3f9b018SEric Dumazet * 799c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 800c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 801c3f9b018SEric Dumazet */ 802c3f9b018SEric Dumazet sock_release_ownership(sk); 803c3f9b018SEric Dumazet 804144d56e9SEric Dumazet if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 8056f458dfbSEric Dumazet tcp_write_timer_handler(sk); 806144d56e9SEric Dumazet __sock_put(sk); 807144d56e9SEric Dumazet } 808144d56e9SEric Dumazet if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 8096f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 810144d56e9SEric Dumazet __sock_put(sk); 811144d56e9SEric Dumazet } 812144d56e9SEric Dumazet if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 8134fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 814144d56e9SEric Dumazet __sock_put(sk); 815144d56e9SEric Dumazet } 81646d3ceabSEric Dumazet } 81746d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 81846d3ceabSEric Dumazet 81946d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 82046d3ceabSEric Dumazet { 82146d3ceabSEric Dumazet int i; 82246d3ceabSEric Dumazet 82346d3ceabSEric Dumazet for_each_possible_cpu(i) { 82446d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 82546d3ceabSEric Dumazet 82646d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 82746d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 82846d3ceabSEric Dumazet tcp_tasklet_func, 82946d3ceabSEric Dumazet (unsigned long)tsq); 83046d3ceabSEric Dumazet } 83146d3ceabSEric Dumazet } 83246d3ceabSEric Dumazet 83346d3ceabSEric Dumazet /* 83446d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 8358e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 83646d3ceabSEric Dumazet * hold qdisc lock. 83746d3ceabSEric Dumazet */ 838d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 83946d3ceabSEric Dumazet { 84046d3ceabSEric Dumazet struct sock *sk = skb->sk; 84146d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 84246d3ceabSEric Dumazet 84346d3ceabSEric Dumazet if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 84446d3ceabSEric Dumazet !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 84546d3ceabSEric Dumazet unsigned long flags; 84646d3ceabSEric Dumazet struct tsq_tasklet *tsq; 84746d3ceabSEric Dumazet 84846d3ceabSEric Dumazet /* Keep a ref on socket. 84946d3ceabSEric Dumazet * This last ref will be released in tcp_tasklet_func() 85046d3ceabSEric Dumazet */ 85146d3ceabSEric Dumazet atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 85246d3ceabSEric Dumazet 85346d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 85446d3ceabSEric Dumazet local_irq_save(flags); 85546d3ceabSEric Dumazet tsq = &__get_cpu_var(tsq_tasklet); 85646d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 85746d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 85846d3ceabSEric Dumazet local_irq_restore(flags); 85946d3ceabSEric Dumazet } else { 86046d3ceabSEric Dumazet sock_wfree(skb); 86146d3ceabSEric Dumazet } 86246d3ceabSEric Dumazet } 86346d3ceabSEric Dumazet 8641da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 8651da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 8661da177e4SLinus Torvalds * transmission and possible later retransmissions. 8671da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 8681da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 8691da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 8701da177e4SLinus Torvalds * device. 8711da177e4SLinus Torvalds * 8721da177e4SLinus Torvalds * We are working here with either a clone of the original 8731da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 8741da177e4SLinus Torvalds */ 875056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 876056834d9SIlpo Järvinen gfp_t gfp_mask) 8771da177e4SLinus Torvalds { 8786687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 879dfb4b9dcSDavid S. Miller struct inet_sock *inet; 880dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 881dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 88233ad798cSAdam Langley struct tcp_out_options opts; 88395c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 884cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 8851da177e4SLinus Torvalds struct tcphdr *th; 8861da177e4SLinus Torvalds int err; 8871da177e4SLinus Torvalds 888dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 889dfb4b9dcSDavid S. Miller 890ccdbb6e9SEric Dumazet if (clone_it) { 891740b0f18SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 892dfb4b9dcSDavid S. Miller 893dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 894dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 895dfb4b9dcSDavid S. Miller else 896dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 897dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 898dfb4b9dcSDavid S. Miller return -ENOBUFS; 899dfb4b9dcSDavid S. Miller } 900dfb4b9dcSDavid S. Miller 901dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 902dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 903dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 90433ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 9051da177e4SLinus Torvalds 9064de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 90733ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 90833ad798cSAdam Langley else 90933ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 91033ad798cSAdam Langley &md5); 91133ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 9121da177e4SLinus Torvalds 913547669d4SEric Dumazet if (tcp_packets_in_flight(tp) == 0) 9146687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 915547669d4SEric Dumazet 916547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 917547669d4SEric Dumazet * another queue. 918547669d4SEric Dumazet */ 919547669d4SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; 9201da177e4SLinus Torvalds 921aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 922aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 92346d3ceabSEric Dumazet 92446d3ceabSEric Dumazet skb_orphan(skb); 92546d3ceabSEric Dumazet skb->sk = sk; 926c9eeec26SEric Dumazet skb->destructor = tcp_wfree; 927b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 92846d3ceabSEric Dumazet atomic_add(skb->truesize, &sk->sk_wmem_alloc); 9291da177e4SLinus Torvalds 9301da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 931aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 932c720c7e8SEric Dumazet th->source = inet->inet_sport; 933c720c7e8SEric Dumazet th->dest = inet->inet_dport; 9341da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 9351da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 936df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 9374de075e0SEric Dumazet tcb->tcp_flags); 938dfb4b9dcSDavid S. Miller 9394de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 9401da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 9411da177e4SLinus Torvalds * is never scaled. 9421da177e4SLinus Torvalds */ 943600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 9441da177e4SLinus Torvalds } else { 9451da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 9461da177e4SLinus Torvalds } 9471da177e4SLinus Torvalds th->check = 0; 9481da177e4SLinus Torvalds th->urg_ptr = 0; 9491da177e4SLinus Torvalds 95033f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 9517691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 9527691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 9531da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 9541da177e4SLinus Torvalds th->urg = 1; 9557691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 9560eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 9577691367dSHerbert Xu th->urg = 1; 9587691367dSHerbert Xu } 9591da177e4SLinus Torvalds } 9601da177e4SLinus Torvalds 961bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 9624de075e0SEric Dumazet if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 963*735d3831SFlorian Westphal tcp_ecn_send(sk, skb, tcp_header_size); 964dfb4b9dcSDavid S. Miller 965cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 966cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 967cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 968a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 969bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 97049a72dfbSAdam Langley md5, sk, NULL, skb); 971cfb6eeb4SYOSHIFUJI Hideaki } 972cfb6eeb4SYOSHIFUJI Hideaki #endif 973cfb6eeb4SYOSHIFUJI Hideaki 974bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 9751da177e4SLinus Torvalds 9764de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 977fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 9781da177e4SLinus Torvalds 9791da177e4SLinus Torvalds if (skb->len != tcp_header_size) 980cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 9811da177e4SLinus Torvalds 982bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 983aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 984aa2ea058STom Herbert tcp_skb_pcount(skb)); 9851da177e4SLinus Torvalds 986cd7d8498SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_segs */ 987cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 988cd7d8498SEric Dumazet 9897faee5c0SEric Dumazet /* Our usage of tstamp should remain private */ 9907faee5c0SEric Dumazet skb->tstamp.tv64 = 0; 991971f10ecSEric Dumazet 992971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 993971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 994971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 995971f10ecSEric Dumazet 996b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 9977faee5c0SEric Dumazet 99883de47cdSHua Zhong if (likely(err <= 0)) 9991da177e4SLinus Torvalds return err; 10001da177e4SLinus Torvalds 10015ee2c941SChristoph Paasch tcp_enter_cwr(sk); 10021da177e4SLinus Torvalds 1003b9df3cb8SGerrit Renker return net_xmit_eval(err); 10041da177e4SLinus Torvalds } 10051da177e4SLinus Torvalds 100667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 10071da177e4SLinus Torvalds * 10081da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 10091da177e4SLinus Torvalds * otherwise socket can stall. 10101da177e4SLinus Torvalds */ 10111da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 10121da177e4SLinus Torvalds { 10131da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10141da177e4SLinus Torvalds 10151da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 10161da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1017f4a775d1SEric Dumazet __skb_header_release(skb); 1018fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 10193ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 10203ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 10211da177e4SLinus Torvalds } 10221da177e4SLinus Torvalds 102367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 1024cf533ea5SEric Dumazet static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1025056834d9SIlpo Järvinen unsigned int mss_now) 1026f6302d1dSDavid S. Miller { 10277b7fc97aSEric Dumazet struct skb_shared_info *shinfo = skb_shinfo(skb); 10287b7fc97aSEric Dumazet 1029c52e2421SEric Dumazet /* Make sure we own this skb before messing gso_size/gso_segs */ 1030c52e2421SEric Dumazet WARN_ON_ONCE(skb_cloned(skb)); 1031c52e2421SEric Dumazet 10328f26fb1cSEric Dumazet if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1033f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1034f6302d1dSDavid S. Miller * non-TSO case. 1035f6302d1dSDavid S. Miller */ 1036cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 10377b7fc97aSEric Dumazet shinfo->gso_size = 0; 10387b7fc97aSEric Dumazet shinfo->gso_type = 0; 1039f6302d1dSDavid S. Miller } else { 1040cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 10417b7fc97aSEric Dumazet shinfo->gso_size = mss_now; 10427b7fc97aSEric Dumazet shinfo->gso_type = sk->sk_gso_type; 10431da177e4SLinus Torvalds } 10441da177e4SLinus Torvalds } 10451da177e4SLinus Torvalds 104691fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 104768f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 104891fed7a1SIlpo Järvinen */ 1049cf533ea5SEric Dumazet static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 105091fed7a1SIlpo Järvinen int decr) 105191fed7a1SIlpo Järvinen { 1052a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1053a47e5a98SIlpo Järvinen 1054dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 105591fed7a1SIlpo Järvinen return; 105691fed7a1SIlpo Järvinen 10576859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 105891fed7a1SIlpo Järvinen tp->fackets_out -= decr; 105991fed7a1SIlpo Järvinen } 106091fed7a1SIlpo Järvinen 1061797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1062797108d1SIlpo Järvinen * tweaks to fix counters 1063797108d1SIlpo Järvinen */ 1064cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1065797108d1SIlpo Järvinen { 1066797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1067797108d1SIlpo Järvinen 1068797108d1SIlpo Järvinen tp->packets_out -= decr; 1069797108d1SIlpo Järvinen 1070797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1071797108d1SIlpo Järvinen tp->sacked_out -= decr; 1072797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1073797108d1SIlpo Järvinen tp->retrans_out -= decr; 1074797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1075797108d1SIlpo Järvinen tp->lost_out -= decr; 1076797108d1SIlpo Järvinen 1077797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1078797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1079797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1080797108d1SIlpo Järvinen 1081797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 1082797108d1SIlpo Järvinen 1083797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1084797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 108552cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1086797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1087797108d1SIlpo Järvinen 1088797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1089797108d1SIlpo Järvinen } 1090797108d1SIlpo Järvinen 1091490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1092490cc7d0SWillem de Bruijn { 1093490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1094490cc7d0SWillem de Bruijn 1095490cc7d0SWillem de Bruijn if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) && 1096490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1097490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1098490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1099490cc7d0SWillem de Bruijn 1100490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1101490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1102490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1103490cc7d0SWillem de Bruijn } 1104490cc7d0SWillem de Bruijn } 1105490cc7d0SWillem de Bruijn 11061da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 11071da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 11081da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 11091da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 11101da177e4SLinus Torvalds */ 1111056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 11126cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 11131da177e4SLinus Torvalds { 11141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11151da177e4SLinus Torvalds struct sk_buff *buff; 11166475be16SDavid S. Miller int nsize, old_factor; 1117b60b49eaSHerbert Xu int nlen; 11189ce01461SIlpo Järvinen u8 flags; 11191da177e4SLinus Torvalds 11202fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 11212fceec13SIlpo Järvinen return -EINVAL; 11226a438bbeSStephen Hemminger 11231da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 11241da177e4SLinus Torvalds if (nsize < 0) 11251da177e4SLinus Torvalds nsize = 0; 11261da177e4SLinus Torvalds 11276cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 11281da177e4SLinus Torvalds return -ENOMEM; 11291da177e4SLinus Torvalds 11301da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 11316cc55e09SOctavian Purdila buff = sk_stream_alloc_skb(sk, nsize, gfp); 11321da177e4SLinus Torvalds if (buff == NULL) 11331da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1134ef5cb973SHerbert Xu 11353ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11363ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1137b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1138b60b49eaSHerbert Xu buff->truesize += nlen; 1139b60b49eaSHerbert Xu skb->truesize -= nlen; 11401da177e4SLinus Torvalds 11411da177e4SLinus Torvalds /* Correct the sequence numbers. */ 11421da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 11431da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 11441da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 11451da177e4SLinus Torvalds 11461da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 11474de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 11484de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 11494de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1150e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 11511da177e4SLinus Torvalds 115284fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 11531da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1154056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1155056834d9SIlpo Järvinen skb_put(buff, nsize), 11561da177e4SLinus Torvalds nsize, 0); 11571da177e4SLinus Torvalds 11581da177e4SLinus Torvalds skb_trim(skb, len); 11591da177e4SLinus Torvalds 11601da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 11611da177e4SLinus Torvalds } else { 116284fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11631da177e4SLinus Torvalds skb_split(skb, buff, len); 11641da177e4SLinus Torvalds } 11651da177e4SLinus Torvalds 11661da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 11671da177e4SLinus Torvalds 1168a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1169490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 11701da177e4SLinus Torvalds 11716475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 11726475be16SDavid S. Miller 11731da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1174846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1175846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 11761da177e4SLinus Torvalds 11776475be16SDavid S. Miller /* If this packet has been sent out already, we must 11786475be16SDavid S. Miller * adjust the various packet counters. 11796475be16SDavid S. Miller */ 1180cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 11816475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 11826475be16SDavid S. Miller tcp_skb_pcount(buff); 11831da177e4SLinus Torvalds 1184797108d1SIlpo Järvinen if (diff) 1185797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 11861da177e4SLinus Torvalds } 11871da177e4SLinus Torvalds 11881da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1189f4a775d1SEric Dumazet __skb_header_release(buff); 1190fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 11911da177e4SLinus Torvalds 11921da177e4SLinus Torvalds return 0; 11931da177e4SLinus Torvalds } 11941da177e4SLinus Torvalds 11951da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 11961da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 11971da177e4SLinus Torvalds * immediately discarded. 11981da177e4SLinus Torvalds */ 1199f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 12001da177e4SLinus Torvalds { 12017b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 12021da177e4SLinus Torvalds int i, k, eat; 12031da177e4SLinus Torvalds 12044fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 12054fa48bf3SEric Dumazet if (eat) { 12064fa48bf3SEric Dumazet __skb_pull(skb, eat); 12074fa48bf3SEric Dumazet len -= eat; 12084fa48bf3SEric Dumazet if (!len) 12094fa48bf3SEric Dumazet return; 12104fa48bf3SEric Dumazet } 12111da177e4SLinus Torvalds eat = len; 12121da177e4SLinus Torvalds k = 0; 12137b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 12147b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 12157b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 12169e903e08SEric Dumazet 12179e903e08SEric Dumazet if (size <= eat) { 1218aff65da0SIan Campbell skb_frag_unref(skb, i); 12199e903e08SEric Dumazet eat -= size; 12201da177e4SLinus Torvalds } else { 12217b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 12221da177e4SLinus Torvalds if (eat) { 12237b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 12247b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 12251da177e4SLinus Torvalds eat = 0; 12261da177e4SLinus Torvalds } 12271da177e4SLinus Torvalds k++; 12281da177e4SLinus Torvalds } 12291da177e4SLinus Torvalds } 12307b7fc97aSEric Dumazet shinfo->nr_frags = k; 12311da177e4SLinus Torvalds 123227a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 12331da177e4SLinus Torvalds skb->data_len -= len; 12341da177e4SLinus Torvalds skb->len = skb->data_len; 12351da177e4SLinus Torvalds } 12361da177e4SLinus Torvalds 123767edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 12381da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 12391da177e4SLinus Torvalds { 124014bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 12411da177e4SLinus Torvalds return -ENOMEM; 12421da177e4SLinus Torvalds 12434fa48bf3SEric Dumazet __pskb_trim_head(skb, len); 12441da177e4SLinus Torvalds 12451da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 124684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 12471da177e4SLinus Torvalds 12481da177e4SLinus Torvalds skb->truesize -= len; 12491da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 12503ab224beSHideo Aoki sk_mem_uncharge(sk, len); 12511da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 12521da177e4SLinus Torvalds 12535b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 12541da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 12555b35e1e6SNeal Cardwell tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 12561da177e4SLinus Torvalds 12571da177e4SLinus Torvalds return 0; 12581da177e4SLinus Torvalds } 12591da177e4SLinus Torvalds 12601b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 12611b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 12625d424d5aSJohn Heffner { 1263cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1264cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 12655d424d5aSJohn Heffner int mss_now; 12665d424d5aSJohn Heffner 12675d424d5aSJohn Heffner /* Calculate base mss without TCP options: 12685d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 12695d424d5aSJohn Heffner */ 12705d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 12715d424d5aSJohn Heffner 127267469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 127367469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 127467469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 127567469601SEric Dumazet 127667469601SEric Dumazet if (dst && dst_allfrag(dst)) 127767469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 127867469601SEric Dumazet } 127967469601SEric Dumazet 12805d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 12815d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 12825d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 12835d424d5aSJohn Heffner 12845d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 12855d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 12865d424d5aSJohn Heffner 12875d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 12885d424d5aSJohn Heffner if (mss_now < 48) 12895d424d5aSJohn Heffner mss_now = 48; 12905d424d5aSJohn Heffner return mss_now; 12915d424d5aSJohn Heffner } 12925d424d5aSJohn Heffner 12931b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 12941b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 12951b63edd6SYuchung Cheng { 12961b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 12971b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 12981b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 12991b63edd6SYuchung Cheng } 13001b63edd6SYuchung Cheng 13015d424d5aSJohn Heffner /* Inverse of above */ 130267469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 13035d424d5aSJohn Heffner { 1304cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1305cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 13065d424d5aSJohn Heffner int mtu; 13075d424d5aSJohn Heffner 13085d424d5aSJohn Heffner mtu = mss + 13095d424d5aSJohn Heffner tp->tcp_header_len + 13105d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 13115d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13125d424d5aSJohn Heffner 131367469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 131467469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 131567469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 131667469601SEric Dumazet 131767469601SEric Dumazet if (dst && dst_allfrag(dst)) 131867469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 131967469601SEric Dumazet } 13205d424d5aSJohn Heffner return mtu; 13215d424d5aSJohn Heffner } 13225d424d5aSJohn Heffner 132367edfef7SAndi Kleen /* MTU probing init per socket */ 13245d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 13255d424d5aSJohn Heffner { 13265d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13275d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 13285d424d5aSJohn Heffner 13295d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 13305d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 13315d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 13325d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 13335d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 13345d424d5aSJohn Heffner } 13354bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 13365d424d5aSJohn Heffner 13371da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 13381da177e4SLinus Torvalds 13391da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 13401da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 13411da177e4SLinus Torvalds 13421da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1343caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 13441da177e4SLinus Torvalds It also does not include TCP options. 13451da177e4SLinus Torvalds 1346d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 13471da177e4SLinus Torvalds 13481da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 13491da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 13501da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 13511da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 13521da177e4SLinus Torvalds 13531da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 13541da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 13551da177e4SLinus Torvalds 1356d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1357d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 13581da177e4SLinus Torvalds */ 13591da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 13601da177e4SLinus Torvalds { 13611da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1362d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 13635d424d5aSJohn Heffner int mss_now; 13641da177e4SLinus Torvalds 13655d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 13665d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 13671da177e4SLinus Torvalds 13685d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1369409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 13701da177e4SLinus Torvalds 13711da177e4SLinus Torvalds /* And store cached results */ 1372d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 13735d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 13745d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1375c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 13761da177e4SLinus Torvalds 13771da177e4SLinus Torvalds return mss_now; 13781da177e4SLinus Torvalds } 13794bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 13801da177e4SLinus Torvalds 13811da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 13821da177e4SLinus Torvalds * and even PMTU discovery events into account. 13831da177e4SLinus Torvalds */ 13840c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 13851da177e4SLinus Torvalds { 1386cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1387cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1388c1b4a7e6SDavid S. Miller u32 mss_now; 138995c96174SEric Dumazet unsigned int header_len; 139033ad798cSAdam Langley struct tcp_out_options opts; 139133ad798cSAdam Langley struct tcp_md5sig_key *md5; 13921da177e4SLinus Torvalds 1393c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1394c1b4a7e6SDavid S. Miller 13951da177e4SLinus Torvalds if (dst) { 13961da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1397d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 13981da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 13991da177e4SLinus Torvalds } 14001da177e4SLinus Torvalds 140133ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 140233ad798cSAdam Langley sizeof(struct tcphdr); 140333ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 140433ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 140533ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 140633ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 140733ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 140833ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 140933ad798cSAdam Langley mss_now -= delta; 141033ad798cSAdam Langley } 1411cfb6eeb4SYOSHIFUJI Hideaki 14121da177e4SLinus Torvalds return mss_now; 14131da177e4SLinus Torvalds } 14141da177e4SLinus Torvalds 141586fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 141686fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 141786fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 141886fd14adSWeiping Pan */ 141986fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1420a762a980SDavid S. Miller { 14219e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1422a762a980SDavid S. Miller 142386fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 142486fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 142586fd14adSWeiping Pan /* Limited by application or receiver window. */ 142686fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 142786fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 142886fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 142986fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 143086fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 143186fd14adSWeiping Pan } 143286fd14adSWeiping Pan tp->snd_cwnd_used = 0; 143386fd14adSWeiping Pan } 143486fd14adSWeiping Pan tp->snd_cwnd_stamp = tcp_time_stamp; 143586fd14adSWeiping Pan } 143686fd14adSWeiping Pan 1437ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1438a762a980SDavid S. Miller { 1439a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1440a762a980SDavid S. Miller 1441ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1442ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1443ca8a2263SNeal Cardwell */ 1444ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1445ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1446ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1447ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1448ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1449ca8a2263SNeal Cardwell } 1450e114a710SEric Dumazet 145124901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1452a762a980SDavid S. Miller /* Network is feed fully. */ 1453a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1454a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1455a762a980SDavid S. Miller } else { 1456a762a980SDavid S. Miller /* Network starves. */ 1457a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1458a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1459a762a980SDavid S. Miller 146015d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 146115d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1462a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1463a762a980SDavid S. Miller } 1464a762a980SDavid S. Miller } 1465a762a980SDavid S. Miller 1466d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1467d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1468d4589926SEric Dumazet { 1469d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1470d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1471d4589926SEric Dumazet } 1472d4589926SEric Dumazet 1473d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1474d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1475d4589926SEric Dumazet * The test is really : 1476d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1477d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1478d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1479d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 14800e3a4803SIlpo Järvinen */ 1481d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1482d4589926SEric Dumazet const struct sk_buff *skb) 1483d4589926SEric Dumazet { 1484d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1485d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1486d4589926SEric Dumazet } 1487d4589926SEric Dumazet 1488d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1489d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1490d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1491d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1492d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1493d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1494d4589926SEric Dumazet */ 1495d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1496cc93fc51SPeter Pan(潘卫平) int nonagle) 1497d4589926SEric Dumazet { 1498d4589926SEric Dumazet return partial && 1499d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1500d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1501d4589926SEric Dumazet } 1502d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1503d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1504d4589926SEric Dumazet const struct sk_buff *skb, 1505d4589926SEric Dumazet unsigned int mss_now, 1506d4589926SEric Dumazet unsigned int max_segs, 1507d4589926SEric Dumazet int nonagle) 1508c1b4a7e6SDavid S. Miller { 1509cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1510d4589926SEric Dumazet u32 partial, needed, window, max_len; 1511c1b4a7e6SDavid S. Miller 151290840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 15131485348dSBen Hutchings max_len = mss_now * max_segs; 15140e3a4803SIlpo Järvinen 15151485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 15161485348dSBen Hutchings return max_len; 15170e3a4803SIlpo Järvinen 15185ea3a748SIlpo Järvinen needed = min(skb->len, window); 15195ea3a748SIlpo Järvinen 15201485348dSBen Hutchings if (max_len <= needed) 15211485348dSBen Hutchings return max_len; 15220e3a4803SIlpo Järvinen 1523d4589926SEric Dumazet partial = needed % mss_now; 1524d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1525d4589926SEric Dumazet * to include this last segment in this skb. 1526d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1527d4589926SEric Dumazet */ 1528cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1529d4589926SEric Dumazet return needed - partial; 1530d4589926SEric Dumazet 1531d4589926SEric Dumazet return needed; 1532c1b4a7e6SDavid S. Miller } 1533c1b4a7e6SDavid S. Miller 1534c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1535c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1536c1b4a7e6SDavid S. Miller */ 1537cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1538cf533ea5SEric Dumazet const struct sk_buff *skb) 1539c1b4a7e6SDavid S. Miller { 1540c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1541c1b4a7e6SDavid S. Miller 1542c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 15434de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 15444de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1545c1b4a7e6SDavid S. Miller return 1; 1546c1b4a7e6SDavid S. Miller 1547c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1548c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1549c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1550c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1551c1b4a7e6SDavid S. Miller 1552c1b4a7e6SDavid S. Miller return 0; 1553c1b4a7e6SDavid S. Miller } 1554c1b4a7e6SDavid S. Miller 1555b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 155667edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1557c1b4a7e6SDavid S. Miller * SKB onto the wire. 1558c1b4a7e6SDavid S. Miller */ 1559cf533ea5SEric Dumazet static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1560056834d9SIlpo Järvinen unsigned int mss_now) 1561c1b4a7e6SDavid S. Miller { 1562c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1563c1b4a7e6SDavid S. Miller 1564f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1565846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1566c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1567c1b4a7e6SDavid S. Miller } 1568c1b4a7e6SDavid S. Miller return tso_segs; 1569c1b4a7e6SDavid S. Miller } 1570c1b4a7e6SDavid S. Miller 1571c1b4a7e6SDavid S. Miller 1572a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1573c1b4a7e6SDavid S. Miller * sent now. 1574c1b4a7e6SDavid S. Miller */ 1575a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1576c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1577c1b4a7e6SDavid S. Miller { 1578c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1579c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1580c1b4a7e6SDavid S. Miller * 1581c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1582c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1583c1b4a7e6SDavid S. Miller */ 1584c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1585a2a385d6SEric Dumazet return true; 1586c1b4a7e6SDavid S. Miller 15879b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 15889b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1589a2a385d6SEric Dumazet return true; 1590c1b4a7e6SDavid S. Miller 1591cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1592a2a385d6SEric Dumazet return true; 1593c1b4a7e6SDavid S. Miller 1594a2a385d6SEric Dumazet return false; 1595c1b4a7e6SDavid S. Miller } 1596c1b4a7e6SDavid S. Miller 1597c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1598a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1599a2a385d6SEric Dumazet const struct sk_buff *skb, 1600056834d9SIlpo Järvinen unsigned int cur_mss) 1601c1b4a7e6SDavid S. Miller { 1602c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1603c1b4a7e6SDavid S. Miller 1604c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1605c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1606c1b4a7e6SDavid S. Miller 160790840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1608c1b4a7e6SDavid S. Miller } 1609c1b4a7e6SDavid S. Miller 1610fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1611c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1612c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1613c1b4a7e6SDavid S. Miller */ 1614cf533ea5SEric Dumazet static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1615c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1616c1b4a7e6SDavid S. Miller { 1617cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1618c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1619c1b4a7e6SDavid S. Miller 1620846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1621c1b4a7e6SDavid S. Miller 1622c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1623c1b4a7e6SDavid S. Miller return 0; 1624c1b4a7e6SDavid S. Miller 1625c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1626056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1627c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1628c1b4a7e6SDavid S. Miller 1629c1b4a7e6SDavid S. Miller return cwnd_quota; 1630c1b4a7e6SDavid S. Miller } 1631c1b4a7e6SDavid S. Miller 163267edfef7SAndi Kleen /* Test if sending is allowed right now. */ 1633a2a385d6SEric Dumazet bool tcp_may_send_now(struct sock *sk) 1634c1b4a7e6SDavid S. Miller { 1635cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1636fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1637c1b4a7e6SDavid S. Miller 1638a02cec21SEric Dumazet return skb && 16390c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1640c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1641a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1642c1b4a7e6SDavid S. Miller } 1643c1b4a7e6SDavid S. Miller 1644c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1645c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1646c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1647c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1648c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1649c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1650c1b4a7e6SDavid S. Miller */ 1651056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1652c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1653c1b4a7e6SDavid S. Miller { 1654c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1655c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 16569ce01461SIlpo Järvinen u8 flags; 1657c1b4a7e6SDavid S. Miller 1658c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1659c8ac3774SHerbert Xu if (skb->len != skb->data_len) 16606cc55e09SOctavian Purdila return tcp_fragment(sk, skb, len, mss_now, gfp); 1661c1b4a7e6SDavid S. Miller 1662c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1663c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1664c1b4a7e6SDavid S. Miller return -ENOMEM; 1665c1b4a7e6SDavid S. Miller 16663ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 16673ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1668b60b49eaSHerbert Xu buff->truesize += nlen; 1669c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1670c1b4a7e6SDavid S. Miller 1671c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1672c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1673c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1674c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1675c1b4a7e6SDavid S. Miller 1676c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 16774de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 16784de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 16794de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1680c1b4a7e6SDavid S. Miller 1681c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1682c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1683c1b4a7e6SDavid S. Miller 168484fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1685c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1686490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1687c1b4a7e6SDavid S. Miller 1688c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1689846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1690846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1691c1b4a7e6SDavid S. Miller 1692c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1693f4a775d1SEric Dumazet __skb_header_release(buff); 1694fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1695c1b4a7e6SDavid S. Miller 1696c1b4a7e6SDavid S. Miller return 0; 1697c1b4a7e6SDavid S. Miller } 1698c1b4a7e6SDavid S. Miller 1699c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1700c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1701c1b4a7e6SDavid S. Miller * 1702c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1703c1b4a7e6SDavid S. Miller */ 1704ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1705ca8a2263SNeal Cardwell bool *is_cwnd_limited) 1706c1b4a7e6SDavid S. Miller { 17079e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 17086687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1709c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1710ad9f4f50SEric Dumazet int win_divisor; 1711c1b4a7e6SDavid S. Miller 17124de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1713ae8064acSJohn Heffner goto send_now; 1714c1b4a7e6SDavid S. Miller 17156687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1716ae8064acSJohn Heffner goto send_now; 1717ae8064acSJohn Heffner 1718ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1719bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1720a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1721ae8064acSJohn Heffner goto send_now; 1722908a75c1SDavid S. Miller 1723c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1724c1b4a7e6SDavid S. Miller 1725056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1726c1b4a7e6SDavid S. Miller 172790840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1728c1b4a7e6SDavid S. Miller 1729c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1730c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1731c1b4a7e6SDavid S. Miller 1732c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1733c1b4a7e6SDavid S. Miller 1734ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 17351485348dSBen Hutchings if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 173695bd09ebSEric Dumazet tp->xmit_size_goal_segs * tp->mss_cache)) 1737ae8064acSJohn Heffner goto send_now; 1738ba244fe9SDavid S. Miller 173962ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 174062ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 174162ad2761SIlpo Järvinen goto send_now; 174262ad2761SIlpo Järvinen 1743ad9f4f50SEric Dumazet win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1744ad9f4f50SEric Dumazet if (win_divisor) { 1745c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1746c1b4a7e6SDavid S. Miller 1747c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1748c1b4a7e6SDavid S. Miller * just use it. 1749c1b4a7e6SDavid S. Miller */ 1750ad9f4f50SEric Dumazet chunk /= win_divisor; 1751c1b4a7e6SDavid S. Miller if (limit >= chunk) 1752ae8064acSJohn Heffner goto send_now; 1753c1b4a7e6SDavid S. Miller } else { 1754c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1755c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1756c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1757c1b4a7e6SDavid S. Miller * then send now. 1758c1b4a7e6SDavid S. Miller */ 17596b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1760ae8064acSJohn Heffner goto send_now; 1761c1b4a7e6SDavid S. Miller } 1762c1b4a7e6SDavid S. Miller 1763f4541d60SEric Dumazet /* Ok, it looks like it is advisable to defer. 1764f4541d60SEric Dumazet * Do not rearm the timer if already set to not break TCP ACK clocking. 1765f4541d60SEric Dumazet */ 1766f4541d60SEric Dumazet if (!tp->tso_deferred) 1767ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1768ae8064acSJohn Heffner 1769ca8a2263SNeal Cardwell if (cong_win < send_win && cong_win < skb->len) 1770ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1771ca8a2263SNeal Cardwell 1772a2a385d6SEric Dumazet return true; 1773ae8064acSJohn Heffner 1774ae8064acSJohn Heffner send_now: 1775ae8064acSJohn Heffner tp->tso_deferred = 0; 1776a2a385d6SEric Dumazet return false; 1777c1b4a7e6SDavid S. Miller } 1778c1b4a7e6SDavid S. Miller 17795d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 178067edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 178167edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 178267edfef7SAndi Kleen * changes resulting in larger path MTUs. 178367edfef7SAndi Kleen * 17845d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 17855d424d5aSJohn Heffner * 1 if a probe was sent, 1786056834d9SIlpo Järvinen * -1 otherwise 1787056834d9SIlpo Järvinen */ 17885d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 17895d424d5aSJohn Heffner { 17905d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 17915d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 17925d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 17935d424d5aSJohn Heffner int len; 17945d424d5aSJohn Heffner int probe_size; 179591cc17c0SIlpo Järvinen int size_needed; 17965d424d5aSJohn Heffner int copy; 17975d424d5aSJohn Heffner int mss_now; 17985d424d5aSJohn Heffner 17995d424d5aSJohn Heffner /* Not currently probing/verifying, 18005d424d5aSJohn Heffner * not in recovery, 18015d424d5aSJohn Heffner * have enough cwnd, and 18025d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 18035d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 18045d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 18055d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 18065d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1807cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 18085d424d5aSJohn Heffner return -1; 18095d424d5aSJohn Heffner 18105d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 18110c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 18125d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 181391cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 18145d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 18155d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 18165d424d5aSJohn Heffner return -1; 18175d424d5aSJohn Heffner } 18185d424d5aSJohn Heffner 18195d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 18207f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 18215d424d5aSJohn Heffner return -1; 18225d424d5aSJohn Heffner 182391cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 18245d424d5aSJohn Heffner return -1; 182590840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 18265d424d5aSJohn Heffner return 0; 18275d424d5aSJohn Heffner 1828d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1829d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1830d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 18315d424d5aSJohn Heffner return -1; 18325d424d5aSJohn Heffner else 18335d424d5aSJohn Heffner return 0; 18345d424d5aSJohn Heffner } 18355d424d5aSJohn Heffner 18365d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 18375d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 18385d424d5aSJohn Heffner return -1; 18393ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 18403ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 18415d424d5aSJohn Heffner 1842fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 18435d424d5aSJohn Heffner 18445d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 18455d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 18464de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 18475d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 18485d424d5aSJohn Heffner nskb->csum = 0; 184984fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 18505d424d5aSJohn Heffner 185150c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 185250c4817eSIlpo Järvinen 18535d424d5aSJohn Heffner len = 0; 1854234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 18555d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 18565d424d5aSJohn Heffner if (nskb->ip_summed) 18575d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 18585d424d5aSJohn Heffner else 18595d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1860056834d9SIlpo Järvinen skb_put(nskb, copy), 1861056834d9SIlpo Järvinen copy, nskb->csum); 18625d424d5aSJohn Heffner 18635d424d5aSJohn Heffner if (skb->len <= copy) { 18645d424d5aSJohn Heffner /* We've eaten all the data from this skb. 18655d424d5aSJohn Heffner * Throw it away. */ 18664de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1867fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 18683ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 18695d424d5aSJohn Heffner } else { 18704de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1871a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 18725d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 18735d424d5aSJohn Heffner skb_pull(skb, copy); 187484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1875056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1876056834d9SIlpo Järvinen skb->len, 0); 18775d424d5aSJohn Heffner } else { 18785d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 18795d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 18805d424d5aSJohn Heffner } 18815d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 18825d424d5aSJohn Heffner } 18835d424d5aSJohn Heffner 18845d424d5aSJohn Heffner len += copy; 1885234b6860SIlpo Järvinen 1886234b6860SIlpo Järvinen if (len >= probe_size) 1887234b6860SIlpo Järvinen break; 18885d424d5aSJohn Heffner } 18895d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 18905d424d5aSJohn Heffner 18915d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 18927faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 18937faee5c0SEric Dumazet */ 18945d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 18955d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 18965d424d5aSJohn Heffner * effectively two packets. */ 18975d424d5aSJohn Heffner tp->snd_cwnd--; 189866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 18995d424d5aSJohn Heffner 19005d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 19010e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 19020e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 19035d424d5aSJohn Heffner 19045d424d5aSJohn Heffner return 1; 19055d424d5aSJohn Heffner } 19065d424d5aSJohn Heffner 19075d424d5aSJohn Heffner return -1; 19085d424d5aSJohn Heffner } 19095d424d5aSJohn Heffner 19101da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 19111da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 19121da177e4SLinus Torvalds * window for us. 19131da177e4SLinus Torvalds * 1914f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1915f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1916f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1917f8269a49SIlpo Järvinen * 19186ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 19196ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 19206ba8a3b1SNandita Dukkipati 1921a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 1922a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 19231da177e4SLinus Torvalds */ 1924a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1925d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 19261da177e4SLinus Torvalds { 19271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 192892df7b51SDavid S. Miller struct sk_buff *skb; 1929c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1930c1b4a7e6SDavid S. Miller int cwnd_quota; 19315d424d5aSJohn Heffner int result; 1932ca8a2263SNeal Cardwell bool is_cwnd_limited = false; 19331da177e4SLinus Torvalds 1934c1b4a7e6SDavid S. Miller sent_pkts = 0; 19355d424d5aSJohn Heffner 1936d5dd9175SIlpo Järvinen if (!push_one) { 19375d424d5aSJohn Heffner /* Do MTU probing. */ 1938d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1939d5dd9175SIlpo Järvinen if (!result) { 1940a2a385d6SEric Dumazet return false; 19415d424d5aSJohn Heffner } else if (result > 0) { 19425d424d5aSJohn Heffner sent_pkts = 1; 19435d424d5aSJohn Heffner } 1944d5dd9175SIlpo Järvinen } 19455d424d5aSJohn Heffner 1946fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1947c8ac3774SHerbert Xu unsigned int limit; 1948c8ac3774SHerbert Xu 1949b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1950c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1951c1b4a7e6SDavid S. Miller 19529d186cacSAndrey Vagin if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 19537faee5c0SEric Dumazet /* "skb_mstamp" is used as a start point for the retransmit timer */ 19547faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 1955ec342325SAndrew Vagin goto repair; /* Skip network transmission */ 19569d186cacSAndrey Vagin } 1957ec342325SAndrew Vagin 1958b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 19596ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 1960ca8a2263SNeal Cardwell is_cwnd_limited = true; 19616ba8a3b1SNandita Dukkipati if (push_one == 2) 19626ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 19636ba8a3b1SNandita Dukkipati cwnd_quota = 1; 19646ba8a3b1SNandita Dukkipati else 1965b68e9f85SHerbert Xu break; 19666ba8a3b1SNandita Dukkipati } 1967b68e9f85SHerbert Xu 1968b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1969b68e9f85SHerbert Xu break; 1970b68e9f85SHerbert Xu 1971c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1972aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1973aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1974aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1975aa93466bSDavid S. Miller break; 1976c1b4a7e6SDavid S. Miller } else { 1977ca8a2263SNeal Cardwell if (!push_one && 1978ca8a2263SNeal Cardwell tcp_tso_should_defer(sk, skb, &is_cwnd_limited)) 1979aa93466bSDavid S. Miller break; 1980c1b4a7e6SDavid S. Miller } 1981aa93466bSDavid S. Miller 1982c9eeec26SEric Dumazet /* TCP Small Queues : 1983c9eeec26SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 1984c9eeec26SEric Dumazet * This allows for : 1985c9eeec26SEric Dumazet * - better RTT estimation and ACK scheduling 1986c9eeec26SEric Dumazet * - faster recovery 1987c9eeec26SEric Dumazet * - high rates 198898e09386SEric Dumazet * Alas, some drivers / subsystems require a fair amount 198998e09386SEric Dumazet * of queued bytes to ensure line rate. 199098e09386SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 199146d3ceabSEric Dumazet */ 199298e09386SEric Dumazet limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes, 199398e09386SEric Dumazet sk->sk_pacing_rate >> 10); 1994c9eeec26SEric Dumazet 1995c9eeec26SEric Dumazet if (atomic_read(&sk->sk_wmem_alloc) > limit) { 199646d3ceabSEric Dumazet set_bit(TSQ_THROTTLED, &tp->tsq_flags); 1997bf06200eSJohn Ogness /* It is possible TX completion already happened 1998bf06200eSJohn Ogness * before we set TSQ_THROTTLED, so we must 1999bf06200eSJohn Ogness * test again the condition. 2000bf06200eSJohn Ogness */ 20014e857c58SPeter Zijlstra smp_mb__after_atomic(); 2002bf06200eSJohn Ogness if (atomic_read(&sk->sk_wmem_alloc) > limit) 200346d3ceabSEric Dumazet break; 200446d3ceabSEric Dumazet } 2005c9eeec26SEric Dumazet 2006c8ac3774SHerbert Xu limit = mss_now; 2007f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 20080e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 20091485348dSBen Hutchings min_t(unsigned int, 20101485348dSBen Hutchings cwnd_quota, 2011d4589926SEric Dumazet sk->sk_gso_max_segs), 2012d4589926SEric Dumazet nonagle); 2013c8ac3774SHerbert Xu 2014c8ac3774SHerbert Xu if (skb->len > limit && 2015c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 20161da177e4SLinus Torvalds break; 20171da177e4SLinus Torvalds 2018d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 20191da177e4SLinus Torvalds break; 20201da177e4SLinus Torvalds 2021ec342325SAndrew Vagin repair: 20221da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 20231da177e4SLinus Torvalds * This call will increment packets_out. 20241da177e4SLinus Torvalds */ 202566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 20261da177e4SLinus Torvalds 20271da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2028a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2029d5dd9175SIlpo Järvinen 2030d5dd9175SIlpo Järvinen if (push_one) 2031d5dd9175SIlpo Järvinen break; 20321da177e4SLinus Torvalds } 20331da177e4SLinus Torvalds 2034aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2035684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2036684bad11SYuchung Cheng tp->prr_out += sent_pkts; 20376ba8a3b1SNandita Dukkipati 20386ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 20396ba8a3b1SNandita Dukkipati if (push_one != 2) 20406ba8a3b1SNandita Dukkipati tcp_schedule_loss_probe(sk); 2041ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2042a2a385d6SEric Dumazet return false; 20431da177e4SLinus Torvalds } 20446ba8a3b1SNandita Dukkipati return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 20456ba8a3b1SNandita Dukkipati } 20466ba8a3b1SNandita Dukkipati 20476ba8a3b1SNandita Dukkipati bool tcp_schedule_loss_probe(struct sock *sk) 20486ba8a3b1SNandita Dukkipati { 20496ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 20506ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 20516ba8a3b1SNandita Dukkipati u32 timeout, tlp_time_stamp, rto_time_stamp; 2052740b0f18SEric Dumazet u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 20536ba8a3b1SNandita Dukkipati 20546ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 20556ba8a3b1SNandita Dukkipati return false; 20566ba8a3b1SNandita Dukkipati /* No consecutive loss probes. */ 20576ba8a3b1SNandita Dukkipati if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 20586ba8a3b1SNandita Dukkipati tcp_rearm_rto(sk); 20596ba8a3b1SNandita Dukkipati return false; 20606ba8a3b1SNandita Dukkipati } 20616ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 20626ba8a3b1SNandita Dukkipati * finishes. 20636ba8a3b1SNandita Dukkipati */ 20646ba8a3b1SNandita Dukkipati if (sk->sk_state == TCP_SYN_RECV) 20656ba8a3b1SNandita Dukkipati return false; 20666ba8a3b1SNandita Dukkipati 20676ba8a3b1SNandita Dukkipati /* TLP is only scheduled when next timer event is RTO. */ 20686ba8a3b1SNandita Dukkipati if (icsk->icsk_pending != ICSK_TIME_RETRANS) 20696ba8a3b1SNandita Dukkipati return false; 20706ba8a3b1SNandita Dukkipati 20716ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 20726ba8a3b1SNandita Dukkipati * in Open state, that are either limited by cwnd or application. 20736ba8a3b1SNandita Dukkipati */ 2074740b0f18SEric Dumazet if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || 20756ba8a3b1SNandita Dukkipati !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 20766ba8a3b1SNandita Dukkipati return false; 20776ba8a3b1SNandita Dukkipati 20786ba8a3b1SNandita Dukkipati if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 20796ba8a3b1SNandita Dukkipati tcp_send_head(sk)) 20806ba8a3b1SNandita Dukkipati return false; 20816ba8a3b1SNandita Dukkipati 20826ba8a3b1SNandita Dukkipati /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 20836ba8a3b1SNandita Dukkipati * for delayed ack when there's one outstanding packet. 20846ba8a3b1SNandita Dukkipati */ 20856ba8a3b1SNandita Dukkipati timeout = rtt << 1; 20866ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 20876ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, 20886ba8a3b1SNandita Dukkipati (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 20896ba8a3b1SNandita Dukkipati timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 20906ba8a3b1SNandita Dukkipati 20916ba8a3b1SNandita Dukkipati /* If RTO is shorter, just schedule TLP in its place. */ 20926ba8a3b1SNandita Dukkipati tlp_time_stamp = tcp_time_stamp + timeout; 20936ba8a3b1SNandita Dukkipati rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 20946ba8a3b1SNandita Dukkipati if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 20956ba8a3b1SNandita Dukkipati s32 delta = rto_time_stamp - tcp_time_stamp; 20966ba8a3b1SNandita Dukkipati if (delta > 0) 20976ba8a3b1SNandita Dukkipati timeout = delta; 20986ba8a3b1SNandita Dukkipati } 20996ba8a3b1SNandita Dukkipati 21006ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 21016ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 21026ba8a3b1SNandita Dukkipati return true; 21036ba8a3b1SNandita Dukkipati } 21046ba8a3b1SNandita Dukkipati 21051f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 21061f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 21071f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 21081f3279aeSEric Dumazet * Note: This is called from BH context only. 21091f3279aeSEric Dumazet */ 21101f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 21111f3279aeSEric Dumazet const struct sk_buff *skb) 21121f3279aeSEric Dumazet { 21131f3279aeSEric Dumazet const struct sk_buff *fclone = skb + 1; 21141f3279aeSEric Dumazet 21151f3279aeSEric Dumazet if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 21161f3279aeSEric Dumazet fclone->fclone == SKB_FCLONE_CLONE)) { 21171f3279aeSEric Dumazet NET_INC_STATS_BH(sock_net(sk), 21181f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 21191f3279aeSEric Dumazet return true; 21201f3279aeSEric Dumazet } 21211f3279aeSEric Dumazet return false; 21221f3279aeSEric Dumazet } 21231f3279aeSEric Dumazet 21246ba8a3b1SNandita Dukkipati /* When probe timeout (PTO) fires, send a new segment if one exists, else 21256ba8a3b1SNandita Dukkipati * retransmit the last segment. 21266ba8a3b1SNandita Dukkipati */ 21276ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 21286ba8a3b1SNandita Dukkipati { 21299b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 21306ba8a3b1SNandita Dukkipati struct sk_buff *skb; 21316ba8a3b1SNandita Dukkipati int pcount; 21326ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 21336ba8a3b1SNandita Dukkipati int err = -1; 21346ba8a3b1SNandita Dukkipati 21356ba8a3b1SNandita Dukkipati if (tcp_send_head(sk) != NULL) { 21366ba8a3b1SNandita Dukkipati err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 21376ba8a3b1SNandita Dukkipati goto rearm_timer; 21386ba8a3b1SNandita Dukkipati } 21396ba8a3b1SNandita Dukkipati 21409b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 21419b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 21429b717a8dSNandita Dukkipati goto rearm_timer; 21439b717a8dSNandita Dukkipati 21446ba8a3b1SNandita Dukkipati /* Retransmit last segment. */ 21456ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 21466ba8a3b1SNandita Dukkipati if (WARN_ON(!skb)) 21476ba8a3b1SNandita Dukkipati goto rearm_timer; 21486ba8a3b1SNandita Dukkipati 21491f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 21501f3279aeSEric Dumazet goto rearm_timer; 21511f3279aeSEric Dumazet 21526ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 21536ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 21546ba8a3b1SNandita Dukkipati goto rearm_timer; 21556ba8a3b1SNandita Dukkipati 21566ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 21576cc55e09SOctavian Purdila if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 21586cc55e09SOctavian Purdila GFP_ATOMIC))) 21596ba8a3b1SNandita Dukkipati goto rearm_timer; 21606ba8a3b1SNandita Dukkipati skb = tcp_write_queue_tail(sk); 21616ba8a3b1SNandita Dukkipati } 21626ba8a3b1SNandita Dukkipati 21636ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 21646ba8a3b1SNandita Dukkipati goto rearm_timer; 21656ba8a3b1SNandita Dukkipati 21666ba8a3b1SNandita Dukkipati err = __tcp_retransmit_skb(sk, skb); 21676ba8a3b1SNandita Dukkipati 21689b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 21699b717a8dSNandita Dukkipati if (likely(!err)) 21709b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 21719b717a8dSNandita Dukkipati 21726ba8a3b1SNandita Dukkipati rearm_timer: 21736ba8a3b1SNandita Dukkipati inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 21746ba8a3b1SNandita Dukkipati inet_csk(sk)->icsk_rto, 21756ba8a3b1SNandita Dukkipati TCP_RTO_MAX); 21766ba8a3b1SNandita Dukkipati 21776ba8a3b1SNandita Dukkipati if (likely(!err)) 21786ba8a3b1SNandita Dukkipati NET_INC_STATS_BH(sock_net(sk), 21796ba8a3b1SNandita Dukkipati LINUX_MIB_TCPLOSSPROBES); 21801da177e4SLinus Torvalds } 21811da177e4SLinus Torvalds 2182a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2183a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2184a762a980SDavid S. Miller * The socket must be locked by the caller. 2185a762a980SDavid S. Miller */ 21869e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 21879e412ba7SIlpo Järvinen int nonagle) 2188a762a980SDavid S. Miller { 2189726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2190726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2191726e07a8SIlpo Järvinen * all will be happy. 2192726e07a8SIlpo Järvinen */ 2193726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2194726e07a8SIlpo Järvinen return; 2195726e07a8SIlpo Järvinen 219699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 219799a1dec7SMel Gorman sk_gfp_atomic(sk, GFP_ATOMIC))) 21989e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2199a762a980SDavid S. Miller } 2200a762a980SDavid S. Miller 2201c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2202c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2203c1b4a7e6SDavid S. Miller */ 2204c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2205c1b4a7e6SDavid S. Miller { 2206fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2207c1b4a7e6SDavid S. Miller 2208c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2209c1b4a7e6SDavid S. Miller 2210d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2211c1b4a7e6SDavid S. Miller } 2212c1b4a7e6SDavid S. Miller 22131da177e4SLinus Torvalds /* This function returns the amount that we can raise the 22141da177e4SLinus Torvalds * usable window based on the following constraints 22151da177e4SLinus Torvalds * 22161da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 22171da177e4SLinus Torvalds * 2. We limit memory per socket 22181da177e4SLinus Torvalds * 22191da177e4SLinus Torvalds * RFC 1122: 22201da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 22211da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 22221da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 22231da177e4SLinus Torvalds * 22241da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 22251da177e4SLinus Torvalds * it at least MSS bytes. 22261da177e4SLinus Torvalds * 22271da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 22281da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 22291da177e4SLinus Torvalds * 22301da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 22311da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 22321da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 22331da177e4SLinus Torvalds * window to always advance by a single byte. 22341da177e4SLinus Torvalds * 22351da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 22361da177e4SLinus Torvalds * then this will not be a problem. 22371da177e4SLinus Torvalds * 22381da177e4SLinus Torvalds * BSD seems to make the following compromise: 22391da177e4SLinus Torvalds * 22401da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 22411da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 22421da177e4SLinus Torvalds * then set the window to 0. 22431da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 22441da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 22451da177e4SLinus Torvalds * and from being larger than the largest representable value. 22461da177e4SLinus Torvalds * 22471da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 22481da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 22491da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 22501da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 22511da177e4SLinus Torvalds * because the pipeline is full. 22521da177e4SLinus Torvalds * 22531da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 22541da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 22551da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 22561da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 22571da177e4SLinus Torvalds * of having a fixed window size at almost all times. 22581da177e4SLinus Torvalds * 22591da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 22601da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 22611da177e4SLinus Torvalds * 22621da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 22631da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 22641da177e4SLinus Torvalds */ 22651da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 22661da177e4SLinus Torvalds { 2267463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 22681da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2269caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 22701da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 22711da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 22721da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 22731da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 22741da177e4SLinus Torvalds */ 2275463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 22761da177e4SLinus Torvalds int free_space = tcp_space(sk); 227786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 227886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 22791da177e4SLinus Torvalds int window; 22801da177e4SLinus Torvalds 22811da177e4SLinus Torvalds if (mss > full_space) 22821da177e4SLinus Torvalds mss = full_space; 22831da177e4SLinus Torvalds 2284b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2285463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 22861da177e4SLinus Torvalds 2287180d8cd9SGlauber Costa if (sk_under_memory_pressure(sk)) 2288056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2289056834d9SIlpo Järvinen 4U * tp->advmss); 22901da177e4SLinus Torvalds 229186c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 229286c1a045SFlorian Westphal * increase it due to wscale. 229386c1a045SFlorian Westphal */ 229486c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 229586c1a045SFlorian Westphal 229686c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 229786c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 229886c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 229986c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 230086c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 230186c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 230286c1a045SFlorian Westphal */ 230386c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 23041da177e4SLinus Torvalds return 0; 23051da177e4SLinus Torvalds } 23061da177e4SLinus Torvalds 23071da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 23081da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 23091da177e4SLinus Torvalds 23101da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 23111da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 23121da177e4SLinus Torvalds */ 23131da177e4SLinus Torvalds window = tp->rcv_wnd; 23141da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 23151da177e4SLinus Torvalds window = free_space; 23161da177e4SLinus Torvalds 23171da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 23181da177e4SLinus Torvalds * Import case: prevent zero window announcement if 23191da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 23201da177e4SLinus Torvalds */ 23211da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 23221da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 23231da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 23241da177e4SLinus Torvalds } else { 23251da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 23261da177e4SLinus Torvalds * Window clamp already applied above. 23271da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 23281da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 23291da177e4SLinus Torvalds * and multiply from happening most of the time. 23301da177e4SLinus Torvalds * We also don't do any window rounding when the free space 23311da177e4SLinus Torvalds * is too small. 23321da177e4SLinus Torvalds */ 23331da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 23341da177e4SLinus Torvalds window = (free_space / mss) * mss; 233584565070SJohn Heffner else if (mss == full_space && 2336b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 233784565070SJohn Heffner window = free_space; 23381da177e4SLinus Torvalds } 23391da177e4SLinus Torvalds 23401da177e4SLinus Torvalds return window; 23411da177e4SLinus Torvalds } 23421da177e4SLinus Torvalds 23434a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 23444a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 23451da177e4SLinus Torvalds { 23461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2347fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2348058dc334SIlpo Järvinen int skb_size, next_skb_size; 23491da177e4SLinus Torvalds 2350058dc334SIlpo Järvinen skb_size = skb->len; 2351058dc334SIlpo Järvinen next_skb_size = next_skb->len; 23521da177e4SLinus Torvalds 2353058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 23541da177e4SLinus Torvalds 23556859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 2356a6963a6bSIlpo Järvinen 2357fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 23581da177e4SLinus Torvalds 2359058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 23601a4e2d09SArnaldo Carvalho de Melo next_skb_size); 23611da177e4SLinus Torvalds 236252d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 236352d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 23641da177e4SLinus Torvalds 236584fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 23661da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 23671da177e4SLinus Torvalds 23681da177e4SLinus Torvalds /* Update sequence range on original skb. */ 23691da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 23701da177e4SLinus Torvalds 2371e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 23724de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 23731da177e4SLinus Torvalds 23741da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 23751da177e4SLinus Torvalds * packet counting does not break. 23761da177e4SLinus Torvalds */ 23774828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2378b7689205SIlpo Järvinen 2379b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2380ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2381ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2382ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2383b7689205SIlpo Järvinen 2384797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2385797108d1SIlpo Järvinen 23863ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 23871da177e4SLinus Torvalds } 23881da177e4SLinus Torvalds 238967edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2390a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 23914a17fc3aSIlpo Järvinen { 23924a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2393a2a385d6SEric Dumazet return false; 23944a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 23954a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 2396a2a385d6SEric Dumazet return false; 23974a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2398a2a385d6SEric Dumazet return false; 23994a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 2400a2a385d6SEric Dumazet return false; 24014a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 24024a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2403a2a385d6SEric Dumazet return false; 24044a17fc3aSIlpo Järvinen 2405a2a385d6SEric Dumazet return true; 24064a17fc3aSIlpo Järvinen } 24074a17fc3aSIlpo Järvinen 240867edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 240967edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 241067edfef7SAndi Kleen */ 24114a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 24124a17fc3aSIlpo Järvinen int space) 24134a17fc3aSIlpo Järvinen { 24144a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 24154a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2416a2a385d6SEric Dumazet bool first = true; 24174a17fc3aSIlpo Järvinen 24184a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 24194a17fc3aSIlpo Järvinen return; 24204de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 24214a17fc3aSIlpo Järvinen return; 24224a17fc3aSIlpo Järvinen 24234a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 24244a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 24254a17fc3aSIlpo Järvinen break; 24264a17fc3aSIlpo Järvinen 24274a17fc3aSIlpo Järvinen space -= skb->len; 24284a17fc3aSIlpo Järvinen 24294a17fc3aSIlpo Järvinen if (first) { 2430a2a385d6SEric Dumazet first = false; 24314a17fc3aSIlpo Järvinen continue; 24324a17fc3aSIlpo Järvinen } 24334a17fc3aSIlpo Järvinen 24344a17fc3aSIlpo Järvinen if (space < 0) 24354a17fc3aSIlpo Järvinen break; 24364a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 24374a17fc3aSIlpo Järvinen * the data in the second 24384a17fc3aSIlpo Järvinen */ 2439a21d4572SEric Dumazet if (skb->len > skb_availroom(to)) 24404a17fc3aSIlpo Järvinen break; 24414a17fc3aSIlpo Järvinen 24424a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 24434a17fc3aSIlpo Järvinen break; 24444a17fc3aSIlpo Järvinen 24454a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 24464a17fc3aSIlpo Järvinen } 24474a17fc3aSIlpo Järvinen } 24484a17fc3aSIlpo Järvinen 24491da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 24501da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 24511da177e4SLinus Torvalds * error occurred which prevented the send. 24521da177e4SLinus Torvalds */ 245393b174adSYuchung Cheng int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 24541da177e4SLinus Torvalds { 24551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24565d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 24577d227cd2SSridhar Samudrala unsigned int cur_mss; 2458c84a5711SYuchung Cheng int err; 24591da177e4SLinus Torvalds 24605d424d5aSJohn Heffner /* Inconslusive MTU probe */ 24615d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 24625d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 24635d424d5aSJohn Heffner } 24645d424d5aSJohn Heffner 24651da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2466caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 24671da177e4SLinus Torvalds */ 24681da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 24691da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 24701da177e4SLinus Torvalds return -EAGAIN; 24711da177e4SLinus Torvalds 24721f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 24731f3279aeSEric Dumazet return -EBUSY; 24741f3279aeSEric Dumazet 24751da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 24761da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 24771da177e4SLinus Torvalds BUG(); 24781da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 24791da177e4SLinus Torvalds return -ENOMEM; 24801da177e4SLinus Torvalds } 24811da177e4SLinus Torvalds 24827d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 24837d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 24847d227cd2SSridhar Samudrala 24850c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 24867d227cd2SSridhar Samudrala 24871da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 24881da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 24891da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 24901da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 24911da177e4SLinus Torvalds */ 24929d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 24939d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 24941da177e4SLinus Torvalds return -EAGAIN; 24951da177e4SLinus Torvalds 24961da177e4SLinus Torvalds if (skb->len > cur_mss) { 24976cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 24981da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 249902276f3cSIlpo Järvinen } else { 25009eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 25019eb9362eSIlpo Järvinen 25029eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 2503c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2504c52e2421SEric Dumazet return -ENOMEM; 250502276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 25069eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 25079eb9362eSIlpo Järvinen } 25081da177e4SLinus Torvalds } 25091da177e4SLinus Torvalds 25101da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 25111da177e4SLinus Torvalds 25121da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 25131da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 25141da177e4SLinus Torvalds */ 25151da177e4SLinus Torvalds 251650bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 251750bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 251850bceae9SThomas Graf * beyond what csum_start can cover. 251950bceae9SThomas Graf */ 252050bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 252150bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 2522117632e6SEric Dumazet struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2523117632e6SEric Dumazet GFP_ATOMIC); 2524c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2525117632e6SEric Dumazet -ENOBUFS; 2526117632e6SEric Dumazet } else { 2527c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2528117632e6SEric Dumazet } 2529c84a5711SYuchung Cheng 2530fc9f3501SEric Dumazet if (likely(!err)) { 2531c84a5711SYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2532fc9f3501SEric Dumazet /* Update global TCP statistics. */ 2533fc9f3501SEric Dumazet TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2534fc9f3501SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2535fc9f3501SEric Dumazet NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2536fc9f3501SEric Dumazet tp->total_retrans++; 2537fc9f3501SEric Dumazet } 2538c84a5711SYuchung Cheng return err; 253993b174adSYuchung Cheng } 254093b174adSYuchung Cheng 254193b174adSYuchung Cheng int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 254293b174adSYuchung Cheng { 254393b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 254493b174adSYuchung Cheng int err = __tcp_retransmit_skb(sk, skb); 25451da177e4SLinus Torvalds 25461da177e4SLinus Torvalds if (err == 0) { 25471da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 25481da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2549e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 25501da177e4SLinus Torvalds } 25511da177e4SLinus Torvalds #endif 2552b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2553b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 25541da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 25551da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 25561da177e4SLinus Torvalds 25571da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 25581da177e4SLinus Torvalds if (!tp->retrans_stamp) 25597faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 25601da177e4SLinus Torvalds 25611da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 25621da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 25631da177e4SLinus Torvalds */ 25641da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 25651f3279aeSEric Dumazet } else if (err != -EBUSY) { 256624ab6becSYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 25671da177e4SLinus Torvalds } 25686e08d5e3SYuchung Cheng 25696e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 25706e08d5e3SYuchung Cheng tp->undo_retrans = 0; 25716e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 25721da177e4SLinus Torvalds return err; 25731da177e4SLinus Torvalds } 25741da177e4SLinus Torvalds 257567edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 257667edfef7SAndi Kleen * window/congestion state. 257767edfef7SAndi Kleen */ 2578a2a385d6SEric Dumazet static bool tcp_can_forward_retransmit(struct sock *sk) 2579b5afe7bcSIlpo Järvinen { 2580b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2581cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 2582b5afe7bcSIlpo Järvinen 2583b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2584b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2585a2a385d6SEric Dumazet return false; 2586b5afe7bcSIlpo Järvinen 2587b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2588b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2589a2a385d6SEric Dumazet return false; 2590b5afe7bcSIlpo Järvinen 2591b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2592b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2593b5afe7bcSIlpo Järvinen * 2594b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2595b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2596b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2597b5afe7bcSIlpo Järvinen */ 2598b5afe7bcSIlpo Järvinen 2599b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2600a2a385d6SEric Dumazet return false; 2601b5afe7bcSIlpo Järvinen 2602a2a385d6SEric Dumazet return true; 2603b5afe7bcSIlpo Järvinen } 2604b5afe7bcSIlpo Järvinen 26051da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 26061da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 26071da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 26081da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 26091da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 26101da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 26111da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 26121da177e4SLinus Torvalds */ 26131da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 26141da177e4SLinus Torvalds { 26156687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 26161da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26171da177e4SLinus Torvalds struct sk_buff *skb; 26180e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2619618d9f25SIlpo Järvinen u32 last_lost; 262061eb55f4SIlpo Järvinen int mib_idx; 26210e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 26226a438bbeSStephen Hemminger 262345e77d31SIlpo Järvinen if (!tp->packets_out) 262445e77d31SIlpo Järvinen return; 262545e77d31SIlpo Järvinen 262608ebd172SIlpo Järvinen if (!tp->lost_out) 262708ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 262808ebd172SIlpo Järvinen 2629618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 26306a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2631618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2632618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2633618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2634618d9f25SIlpo Järvinen } else { 2635fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2636618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2637618d9f25SIlpo Järvinen } 26381da177e4SLinus Torvalds 2639fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 26401da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 26411da177e4SLinus Torvalds 2642fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2643fe067e8aSDavid S. Miller break; 26446a438bbeSStephen Hemminger /* we could do better than to assign each time */ 26450e1c54c2SIlpo Järvinen if (hole == NULL) 26466a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 26476a438bbeSStephen Hemminger 26481da177e4SLinus Torvalds /* Assume this retransmit will generate 26491da177e4SLinus Torvalds * only one packet for congestion window 26501da177e4SLinus Torvalds * calculation purposes. This works because 26511da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 26521da177e4SLinus Torvalds * packet to be MSS sized and all the 26531da177e4SLinus Torvalds * packet counting works out. 26541da177e4SLinus Torvalds */ 26551da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 26561da177e4SLinus Torvalds return; 26570e1c54c2SIlpo Järvinen 26580e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 26590e1c54c2SIlpo Järvinen begin_fwd: 26600e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2661006f582cSIlpo Järvinen break; 26620e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 26630e1c54c2SIlpo Järvinen 26640e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2665618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 26660e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 26670e1c54c2SIlpo Järvinen break; 26680e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 26690e1c54c2SIlpo Järvinen if (hole != NULL) { 26700e1c54c2SIlpo Järvinen skb = hole; 26710e1c54c2SIlpo Järvinen hole = NULL; 26720e1c54c2SIlpo Järvinen } 26730e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 26740e1c54c2SIlpo Järvinen goto begin_fwd; 26750e1c54c2SIlpo Järvinen 26760e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2677ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 26780e1c54c2SIlpo Järvinen hole = skb; 267961eb55f4SIlpo Järvinen continue; 26801da177e4SLinus Torvalds 26810e1c54c2SIlpo Järvinen } else { 2682618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 26830e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 26840e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 26850e1c54c2SIlpo Järvinen else 26860e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 26870e1c54c2SIlpo Järvinen } 26880e1c54c2SIlpo Järvinen 26890e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 269061eb55f4SIlpo Järvinen continue; 269140b215e5SPavel Emelyanov 269224ab6becSYuchung Cheng if (tcp_retransmit_skb(sk, skb)) 26931da177e4SLinus Torvalds return; 269424ab6becSYuchung Cheng 2695de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 26961da177e4SLinus Torvalds 2697684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2698a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 2699a262f0cdSNandita Dukkipati 2700fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2701463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 27023f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 27033f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 27041da177e4SLinus Torvalds } 27051da177e4SLinus Torvalds } 27061da177e4SLinus Torvalds 27071da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 27081da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 27091da177e4SLinus Torvalds */ 27101da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 27111da177e4SLinus Torvalds { 27121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2713fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 27141da177e4SLinus Torvalds int mss_now; 27151da177e4SLinus Torvalds 27161da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 27171da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 27181da177e4SLinus Torvalds * and IP options. 27191da177e4SLinus Torvalds */ 27200c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 27211da177e4SLinus Torvalds 2722fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 27234de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 27241da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 27251da177e4SLinus Torvalds tp->write_seq++; 27261da177e4SLinus Torvalds } else { 27271da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 27281da177e4SLinus Torvalds for (;;) { 2729aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2730aa133076SWu Fengguang sk->sk_allocation); 27311da177e4SLinus Torvalds if (skb) 27321da177e4SLinus Torvalds break; 27331da177e4SLinus Torvalds yield(); 27341da177e4SLinus Torvalds } 27351da177e4SLinus Torvalds 27361da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27371da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 27381da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2739e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2740a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 27411da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 27421da177e4SLinus Torvalds } 27439e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 27441da177e4SLinus Torvalds } 27451da177e4SLinus Torvalds 27461da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 27471da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 27481da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 274965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 27501da177e4SLinus Torvalds */ 2751dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 27521da177e4SLinus Torvalds { 27531da177e4SLinus Torvalds struct sk_buff *skb; 27541da177e4SLinus Torvalds 27551da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 27561da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 27571da177e4SLinus Torvalds if (!skb) { 27584e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 27591da177e4SLinus Torvalds return; 27601da177e4SLinus Torvalds } 27611da177e4SLinus Torvalds 27621da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27631da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2764e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2765a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 27661da177e4SLinus Torvalds /* Send it off. */ 2767dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 27684e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 276926af65cbSSridhar Samudrala 277081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 27711da177e4SLinus Torvalds } 27721da177e4SLinus Torvalds 277367edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 277467edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 27751da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 27761da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 27771da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 27781da177e4SLinus Torvalds */ 27791da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 27801da177e4SLinus Torvalds { 27811da177e4SLinus Torvalds struct sk_buff *skb; 27821da177e4SLinus Torvalds 2783fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 27844de075e0SEric Dumazet if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 278591df42beSJoe Perches pr_debug("%s: wrong queue state\n", __func__); 27861da177e4SLinus Torvalds return -EFAULT; 27871da177e4SLinus Torvalds } 27884de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 27891da177e4SLinus Torvalds if (skb_cloned(skb)) { 27901da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 27911da177e4SLinus Torvalds if (nskb == NULL) 27921da177e4SLinus Torvalds return -ENOMEM; 2793fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 2794f4a775d1SEric Dumazet __skb_header_release(nskb); 2795fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 27963ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 27973ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 27983ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 27991da177e4SLinus Torvalds skb = nskb; 28001da177e4SLinus Torvalds } 28011da177e4SLinus Torvalds 28024de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2803*735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 28041da177e4SLinus Torvalds } 2805dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 28061da177e4SLinus Torvalds } 28071da177e4SLinus Torvalds 28084aea39c1SEric Dumazet /** 28094aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 28104aea39c1SEric Dumazet * sk: listener socket 28114aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 28124aea39c1SEric Dumazet * req: request_sock pointer 28134aea39c1SEric Dumazet * 28144aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 28154aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 28164aea39c1SEric Dumazet */ 28171da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2818e6b4d113SWilliam Allen Simpson struct request_sock *req, 28198336886fSJerry Chu struct tcp_fastopen_cookie *foc) 28201da177e4SLinus Torvalds { 2821bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 28222e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 28231da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28241da177e4SLinus Torvalds struct tcphdr *th; 28251da177e4SLinus Torvalds struct sk_buff *skb; 2826cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2827bd0388aeSWilliam Allen Simpson int tcp_header_size; 2828f5fff5dcSTom Quetchenbach int mss; 28291da177e4SLinus Torvalds 2830a0b8486cSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); 28314aea39c1SEric Dumazet if (unlikely(!skb)) { 28324aea39c1SEric Dumazet dst_release(dst); 28331da177e4SLinus Torvalds return NULL; 28344aea39c1SEric Dumazet } 28351da177e4SLinus Torvalds /* Reserve space for headers. */ 28361da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 28371da177e4SLinus Torvalds 28384aea39c1SEric Dumazet skb_dst_set(skb, dst); 2839ca10b9e9SEric Dumazet security_skb_owned_by(skb, sk); 28401da177e4SLinus Torvalds 28410dbaee3bSDavid S. Miller mss = dst_metric_advmss(dst); 2842f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2843f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2844f5fff5dcSTom Quetchenbach 284533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 28468b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 28478b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 28487faee5c0SEric Dumazet skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); 28498b5f12d0SFlorian Westphal else 28508b5f12d0SFlorian Westphal #endif 28517faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 28521a2c6181SChristoph Paasch tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 28531a2c6181SChristoph Paasch foc) + sizeof(*th); 285433ad798cSAdam Langley 2855aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2856aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 28571da177e4SLinus Torvalds 2858aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 28591da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 28601da177e4SLinus Torvalds th->syn = 1; 28611da177e4SLinus Torvalds th->ack = 1; 2862*735d3831SFlorian Westphal tcp_ecn_make_synack(req, th, sk); 2863b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 2864634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 2865e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2866e870a8efSIlpo Järvinen * not even correctly set) 2867e870a8efSIlpo Järvinen */ 2868e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2869a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 28704957faadSWilliam Allen Simpson 28711da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 28728336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 28738336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 28741da177e4SLinus Torvalds 28751da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2876600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2877bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 28781da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2879a0b8486cSEric Dumazet TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 2880cfb6eeb4SYOSHIFUJI Hideaki 2881cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2882cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2883cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2884bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 288549a72dfbSAdam Langley md5, NULL, req, skb); 2886cfb6eeb4SYOSHIFUJI Hideaki } 2887cfb6eeb4SYOSHIFUJI Hideaki #endif 2888cfb6eeb4SYOSHIFUJI Hideaki 28891da177e4SLinus Torvalds return skb; 28901da177e4SLinus Torvalds } 28914bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 28921da177e4SLinus Torvalds 289367edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 2894f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 28951da177e4SLinus Torvalds { 2896cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 28971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28981da177e4SLinus Torvalds __u8 rcv_wscale; 28991da177e4SLinus Torvalds 29001da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 29011da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 29021da177e4SLinus Torvalds */ 29031da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2904bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 29051da177e4SLinus Torvalds 2906cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2907cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2908cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2909cfb6eeb4SYOSHIFUJI Hideaki #endif 2910cfb6eeb4SYOSHIFUJI Hideaki 29111da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 29121da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 29131da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 29141da177e4SLinus Torvalds tp->max_window = 0; 29155d424d5aSJohn Heffner tcp_mtup_init(sk); 29161da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 29171da177e4SLinus Torvalds 29181da177e4SLinus Torvalds if (!tp->window_clamp) 29191da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 29200dbaee3bSDavid S. Miller tp->advmss = dst_metric_advmss(dst); 2921f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2922f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2923f5fff5dcSTom Quetchenbach 29241da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 29251da177e4SLinus Torvalds 2926e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2927e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2928e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2929e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2930e88c64f0SHagen Paul Pfeifer 29311da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 29321da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 29331da177e4SLinus Torvalds &tp->rcv_wnd, 29341da177e4SLinus Torvalds &tp->window_clamp, 2935bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 293631d12926Slaurent chavey &rcv_wscale, 293731d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 29381da177e4SLinus Torvalds 29391da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 29401da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 29411da177e4SLinus Torvalds 29421da177e4SLinus Torvalds sk->sk_err = 0; 29431da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 29441da177e4SLinus Torvalds tp->snd_wnd = 0; 2945ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 29461da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 29471da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 294833f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 2949370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 2950ee995283SPavel Emelyanov 2951ee995283SPavel Emelyanov if (likely(!tp->repair)) 29521da177e4SLinus Torvalds tp->rcv_nxt = 0; 2953c7781a6eSAndrew Vagin else 2954c7781a6eSAndrew Vagin tp->rcv_tstamp = tcp_time_stamp; 2955ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 2956ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 29571da177e4SLinus Torvalds 2958463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2959463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 29601da177e4SLinus Torvalds tcp_clear_retrans(tp); 29611da177e4SLinus Torvalds } 29621da177e4SLinus Torvalds 2963783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2964783237e8SYuchung Cheng { 2965783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2966783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2967783237e8SYuchung Cheng 2968783237e8SYuchung Cheng tcb->end_seq += skb->len; 2969f4a775d1SEric Dumazet __skb_header_release(skb); 2970783237e8SYuchung Cheng __tcp_add_write_queue_tail(sk, skb); 2971783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 2972783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 2973783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 2974783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 2975783237e8SYuchung Cheng } 2976783237e8SYuchung Cheng 2977783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2978783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 2979783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2980783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 2981783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 2982783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 2983783237e8SYuchung Cheng */ 2984783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2985783237e8SYuchung Cheng { 2986783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2987783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 2988aab48743SYuchung Cheng int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2989783237e8SYuchung Cheng struct sk_buff *syn_data = NULL, *data; 2990aab48743SYuchung Cheng unsigned long last_syn_loss = 0; 2991783237e8SYuchung Cheng 299267da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2993aab48743SYuchung Cheng tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2994aab48743SYuchung Cheng &syn_loss, &last_syn_loss); 2995aab48743SYuchung Cheng /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2996aab48743SYuchung Cheng if (syn_loss > 1 && 2997aab48743SYuchung Cheng time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2998aab48743SYuchung Cheng fo->cookie.len = -1; 2999aab48743SYuchung Cheng goto fallback; 3000aab48743SYuchung Cheng } 3001aab48743SYuchung Cheng 300267da22d2SYuchung Cheng if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 300367da22d2SYuchung Cheng fo->cookie.len = -1; 300467da22d2SYuchung Cheng else if (fo->cookie.len <= 0) 3005783237e8SYuchung Cheng goto fallback; 3006783237e8SYuchung Cheng 3007783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3008783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3009783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3010783237e8SYuchung Cheng */ 3011783237e8SYuchung Cheng if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 3012783237e8SYuchung Cheng tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 30131b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3014783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3015783237e8SYuchung Cheng 3016f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3017f5ddcbbbSEric Dumazet 3018f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3019f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3020f5ddcbbbSEric Dumazet 3021f5ddcbbbSEric Dumazet syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space, 3022783237e8SYuchung Cheng sk->sk_allocation); 3023783237e8SYuchung Cheng if (syn_data == NULL) 3024783237e8SYuchung Cheng goto fallback; 3025783237e8SYuchung Cheng 3026783237e8SYuchung Cheng for (i = 0; i < iovlen && syn_data->len < space; ++i) { 3027783237e8SYuchung Cheng struct iovec *iov = &fo->data->msg_iov[i]; 3028783237e8SYuchung Cheng unsigned char __user *from = iov->iov_base; 3029783237e8SYuchung Cheng int len = iov->iov_len; 3030783237e8SYuchung Cheng 3031783237e8SYuchung Cheng if (syn_data->len + len > space) 3032783237e8SYuchung Cheng len = space - syn_data->len; 3033783237e8SYuchung Cheng else if (i + 1 == iovlen) 3034783237e8SYuchung Cheng /* No more data pending in inet_wait_for_connect() */ 3035783237e8SYuchung Cheng fo->data = NULL; 3036783237e8SYuchung Cheng 3037783237e8SYuchung Cheng if (skb_add_data(syn_data, from, len)) 3038783237e8SYuchung Cheng goto fallback; 3039783237e8SYuchung Cheng } 3040783237e8SYuchung Cheng 3041783237e8SYuchung Cheng /* Queue a data-only packet after the regular SYN for retransmission */ 3042783237e8SYuchung Cheng data = pskb_copy(syn_data, sk->sk_allocation); 3043783237e8SYuchung Cheng if (data == NULL) 3044783237e8SYuchung Cheng goto fallback; 3045783237e8SYuchung Cheng TCP_SKB_CB(data)->seq++; 3046783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 3047783237e8SYuchung Cheng TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 3048783237e8SYuchung Cheng tcp_connect_queue_skb(sk, data); 3049783237e8SYuchung Cheng fo->copied = data->len; 3050783237e8SYuchung Cheng 3051431a9124SEric Dumazet /* syn_data is about to be sent, we need to take current time stamps 3052431a9124SEric Dumazet * for the packets that are in write queue : SYN packet and DATA 3053431a9124SEric Dumazet */ 3054431a9124SEric Dumazet skb_mstamp_get(&syn->skb_mstamp); 3055431a9124SEric Dumazet data->skb_mstamp = syn->skb_mstamp; 3056431a9124SEric Dumazet 3057783237e8SYuchung Cheng if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 305867da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 3059f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3060783237e8SYuchung Cheng goto done; 3061783237e8SYuchung Cheng } 3062783237e8SYuchung Cheng syn_data = NULL; 3063783237e8SYuchung Cheng 3064783237e8SYuchung Cheng fallback: 3065783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3066783237e8SYuchung Cheng if (fo->cookie.len > 0) 3067783237e8SYuchung Cheng fo->cookie.len = 0; 3068783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3069783237e8SYuchung Cheng if (err) 3070783237e8SYuchung Cheng tp->syn_fastopen = 0; 3071783237e8SYuchung Cheng kfree_skb(syn_data); 3072783237e8SYuchung Cheng done: 3073783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3074783237e8SYuchung Cheng return err; 3075783237e8SYuchung Cheng } 3076783237e8SYuchung Cheng 307767edfef7SAndi Kleen /* Build a SYN and send it off. */ 30781da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 30791da177e4SLinus Torvalds { 30801da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30811da177e4SLinus Torvalds struct sk_buff *buff; 3082ee586811SEric Paris int err; 30831da177e4SLinus Torvalds 30841da177e4SLinus Torvalds tcp_connect_init(sk); 30851da177e4SLinus Torvalds 30862b916477SAndrey Vagin if (unlikely(tp->repair)) { 30872b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 30882b916477SAndrey Vagin return 0; 30892b916477SAndrey Vagin } 30902b916477SAndrey Vagin 3091d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 30921da177e4SLinus Torvalds if (unlikely(buff == NULL)) 30931da177e4SLinus Torvalds return -ENOBUFS; 30941da177e4SLinus Torvalds 30951da177e4SLinus Torvalds /* Reserve space for headers. */ 30961da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 30971da177e4SLinus Torvalds 3098a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 30997faee5c0SEric Dumazet tp->retrans_stamp = tcp_time_stamp; 3100783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3101*735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 31021da177e4SLinus Torvalds 3103783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3104783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3105783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3106ee586811SEric Paris if (err == -ECONNREFUSED) 3107ee586811SEric Paris return err; 3108bd37a088SWei Yongjun 3109bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3110bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3111bd37a088SWei Yongjun */ 3112bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3113bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 311481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 31151da177e4SLinus Torvalds 31161da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 31173f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 31183f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 31191da177e4SLinus Torvalds return 0; 31201da177e4SLinus Torvalds } 31214bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 31221da177e4SLinus Torvalds 31231da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 31241da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 31251da177e4SLinus Torvalds * for details. 31261da177e4SLinus Torvalds */ 31271da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 31281da177e4SLinus Torvalds { 3129463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3130463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 31311da177e4SLinus Torvalds unsigned long timeout; 31321da177e4SLinus Torvalds 31339890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); 31349890092eSFlorian Westphal 31351da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3136463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 31371da177e4SLinus Torvalds int max_ato = HZ / 2; 31381da177e4SLinus Torvalds 3139056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 3140056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 31411da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 31421da177e4SLinus Torvalds 31431da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 31441da177e4SLinus Torvalds 31451da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3146463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 31471da177e4SLinus Torvalds * directly. 31481da177e4SLinus Torvalds */ 3149740b0f18SEric Dumazet if (tp->srtt_us) { 3150740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3151740b0f18SEric Dumazet TCP_DELACK_MIN); 31521da177e4SLinus Torvalds 31531da177e4SLinus Torvalds if (rtt < max_ato) 31541da177e4SLinus Torvalds max_ato = rtt; 31551da177e4SLinus Torvalds } 31561da177e4SLinus Torvalds 31571da177e4SLinus Torvalds ato = min(ato, max_ato); 31581da177e4SLinus Torvalds } 31591da177e4SLinus Torvalds 31601da177e4SLinus Torvalds /* Stay within the limit we were given */ 31611da177e4SLinus Torvalds timeout = jiffies + ato; 31621da177e4SLinus Torvalds 31631da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3164463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 31651da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 31661da177e4SLinus Torvalds * send ACK now. 31671da177e4SLinus Torvalds */ 3168463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3169463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 31701da177e4SLinus Torvalds tcp_send_ack(sk); 31711da177e4SLinus Torvalds return; 31721da177e4SLinus Torvalds } 31731da177e4SLinus Torvalds 3174463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3175463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 31761da177e4SLinus Torvalds } 3177463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3178463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3179463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 31801da177e4SLinus Torvalds } 31811da177e4SLinus Torvalds 31821da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 31831da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 31841da177e4SLinus Torvalds { 31851da177e4SLinus Torvalds struct sk_buff *buff; 31861da177e4SLinus Torvalds 3187058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3188058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3189058dc334SIlpo Järvinen return; 3190058dc334SIlpo Järvinen 31919890092eSFlorian Westphal tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); 31929890092eSFlorian Westphal 31931da177e4SLinus Torvalds /* We are not putting this on the write queue, so 31941da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 31951da177e4SLinus Torvalds * sock. 31961da177e4SLinus Torvalds */ 319799a1dec7SMel Gorman buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 31981da177e4SLinus Torvalds if (buff == NULL) { 3199463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3200463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 32013f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 32023f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 32031da177e4SLinus Torvalds return; 32041da177e4SLinus Torvalds } 32051da177e4SLinus Torvalds 32061da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 32071da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3208a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 32091da177e4SLinus Torvalds 32101da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 32117faee5c0SEric Dumazet skb_mstamp_get(&buff->skb_mstamp); 321299a1dec7SMel Gorman tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 32131da177e4SLinus Torvalds } 3214e3118e83SDaniel Borkmann EXPORT_SYMBOL_GPL(tcp_send_ack); 32151da177e4SLinus Torvalds 32161da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 32171da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 32181da177e4SLinus Torvalds * 32191da177e4SLinus Torvalds * Question: what should we make while urgent mode? 32201da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 32211da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 32221da177e4SLinus Torvalds * 32231da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 32241da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 32251da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 32261da177e4SLinus Torvalds */ 32271da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 32281da177e4SLinus Torvalds { 32291da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32301da177e4SLinus Torvalds struct sk_buff *skb; 32311da177e4SLinus Torvalds 32321da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 323399a1dec7SMel Gorman skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 32341da177e4SLinus Torvalds if (skb == NULL) 32351da177e4SLinus Torvalds return -1; 32361da177e4SLinus Torvalds 32371da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 32381da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32391da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 32401da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 32411da177e4SLinus Torvalds * send it. 32421da177e4SLinus Torvalds */ 3243a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 32447faee5c0SEric Dumazet skb_mstamp_get(&skb->skb_mstamp); 3245dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 32461da177e4SLinus Torvalds } 32471da177e4SLinus Torvalds 3248ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3249ee995283SPavel Emelyanov { 3250ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3251ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3252ee995283SPavel Emelyanov tcp_xmit_probe_skb(sk, 0); 3253ee995283SPavel Emelyanov } 3254ee995283SPavel Emelyanov } 3255ee995283SPavel Emelyanov 325667edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 32571da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 32581da177e4SLinus Torvalds { 32591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 32601da177e4SLinus Torvalds struct sk_buff *skb; 32611da177e4SLinus Torvalds 3262058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3263058dc334SIlpo Järvinen return -1; 3264058dc334SIlpo Järvinen 3265fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 326690840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 32671da177e4SLinus Torvalds int err; 32680c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 326990840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 32701da177e4SLinus Torvalds 32711da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 32721da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 32731da177e4SLinus Torvalds 32741da177e4SLinus Torvalds /* We are probing the opening of a window 32751da177e4SLinus Torvalds * but the window size is != 0 32761da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 32771da177e4SLinus Torvalds */ 32781da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 32791da177e4SLinus Torvalds skb->len > mss) { 32801da177e4SLinus Torvalds seg_size = min(seg_size, mss); 32814de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 32826cc55e09SOctavian Purdila if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 32831da177e4SLinus Torvalds return -1; 32841da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 3285846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 32861da177e4SLinus Torvalds 32874de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3288dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 328966f5fe62SIlpo Järvinen if (!err) 329066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 32911da177e4SLinus Torvalds return err; 32921da177e4SLinus Torvalds } else { 329333f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 32944828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 32951da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 32961da177e4SLinus Torvalds } 32971da177e4SLinus Torvalds } 32981da177e4SLinus Torvalds 32991da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 33001da177e4SLinus Torvalds * a partial packet else a zero probe. 33011da177e4SLinus Torvalds */ 33021da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 33031da177e4SLinus Torvalds { 3304463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 33051da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3306fcdd1cf4SEric Dumazet unsigned long probe_max; 33071da177e4SLinus Torvalds int err; 33081da177e4SLinus Torvalds 33091da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 33101da177e4SLinus Torvalds 3311fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 33121da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 33136687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3314463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 33151da177e4SLinus Torvalds return; 33161da177e4SLinus Torvalds } 33171da177e4SLinus Torvalds 33181da177e4SLinus Torvalds if (err <= 0) { 3319463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 3320463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 33216687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 3322fcdd1cf4SEric Dumazet probe_max = TCP_RTO_MAX; 33231da177e4SLinus Torvalds } else { 33241da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 33256687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 33261da177e4SLinus Torvalds * Let local senders to fight for local resources. 33271da177e4SLinus Torvalds * 33281da177e4SLinus Torvalds * Use accumulated backoff yet. 33291da177e4SLinus Torvalds */ 33306687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 33316687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 3332fcdd1cf4SEric Dumazet probe_max = TCP_RESOURCE_PROBE_INTERVAL; 33331da177e4SLinus Torvalds } 3334fcdd1cf4SEric Dumazet inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3335fcdd1cf4SEric Dumazet inet_csk_rto_backoff(icsk, probe_max), 3336fcdd1cf4SEric Dumazet TCP_RTO_MAX); 33371da177e4SLinus Torvalds } 33385db92c99SOctavian Purdila 33395db92c99SOctavian Purdila int tcp_rtx_synack(struct sock *sk, struct request_sock *req) 33405db92c99SOctavian Purdila { 33415db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 33425db92c99SOctavian Purdila struct flowi fl; 33435db92c99SOctavian Purdila int res; 33445db92c99SOctavian Purdila 33455db92c99SOctavian Purdila res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL); 33465db92c99SOctavian Purdila if (!res) { 33475db92c99SOctavian Purdila TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS); 33485db92c99SOctavian Purdila NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 33495db92c99SOctavian Purdila } 33505db92c99SOctavian Purdila return res; 33515db92c99SOctavian Purdila } 33525db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3353